Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
19
99
source_id
stringclasses
9 values
source_dataset
stringclasses
9 values
domain
stringclasses
9 values
template_id
stringclasses
11 values
question
stringlengths
951
2.44k
answer_format
stringclasses
7 values
answer
stringlengths
1
7.14k
unit
stringclasses
10 values
tolerance
stringlengths
27
133
item_dir
stringlengths
24
104
input_files
listlengths
2
148
input_bytes
int64
2.08M
140M
n_materials
int64
11
775
gt_type
stringclasses
4 values
gt_source
stringlengths
245
748
provenance
stringlengths
86
575
make_gt
stringlengths
35
115
prep_notes
stringlengths
292
585
expected_approach
stringlengths
150
330
why_multi_look
stringlengths
259
655
shortcut_risk
stringlengths
184
398
specialist_check
stringlengths
184
553
⌀
opus_scores
listlengths
3
3
opus_answers
stringlengths
10
1.52k
specialist_score
float64
0
0.71
⌀
blind_correct
unknown
audit_category
stringclasses
1 value
audit_evidence
stringlengths
605
2.39k
why_agent_failed
stringlengths
73
589
audit_action
stringlengths
30
334
ap-v01__ht-091bcb0c
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [2881.0, 3600.0, 4320.0, 5040.0, 5760.0, 6482.0, 7200.0, 7920.0, 10182.0, 10810.0, 11520.0, 12240.0, 12601.0, 12961.0, 13320.0, 13680.0, 14041.0, 14400.0, 14762.0, 15120.0, 16200.0, 18360.0, 19080.0, 19800.0, 20521.0, 21961.0, 22681.0, 23400.0, 24121.0], "y": [2.61, 2.83, 2.91, 3.13, 3.19, 3.51, 3.64, 3.85, 4.28,...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-091bcb0c
[ "items/ap-v01/ht-091bcb0c/input/frames.csv", "items/ap-v01/ht-091bcb0c/input/c2/0001.png", "items/ap-v01/ht-091bcb0c/input/c2/0002.png", "items/ap-v01/ht-091bcb0c/input/c2/0003.png", "items/ap-v01/ht-091bcb0c/input/c2/0004.png", "items/ap-v01/ht-091bcb0c/input/c2/0005.png", "items/ap-v01/ht-091bcb0c/inp...
48,621,733
49
Expert annotation
CDAW SOHO/LASCO CME catalogue 2016/04/08 23:24:05 CPA 86 W 55 V 185.0 acc 1.7 MPA 104 | ; expert height-time file 20160408.232405.w055n.v0185.p104g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2016_04/yht/20160408.232405.w055n.v0185.p104g.yht, QUALITY_INDEX 3, observer michalek), all 29 C2/C3 points, t is se...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/160408/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/160408/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/160409/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/160409/c3/ (level-0.5 FITS)", "ht...
items/ap-v01/ht-091bcb0c/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard the FITS header, provide on...
After exposure normalization, compute running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and measured at its farthest point, while also being distinguished from other concurrent CMEs, streamers and residual structures; the entire sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or no count); constant-speed extrapolation at a typical speed earns only partial credit; single-value speed tasks were already guessed from priors in older tasks, so this source uses only curves, combinations, and multi-target lists. For the difficulty pr...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for task screening.
[ 0.517, 0.759, 0.103 ]
["16-point csv, t 2881–15120 s, y 2.30–6.05", "29-point csv, t 5040–25562 s, y 2.69–9.90", "24-point csv, t 2881–21961 s, y 2.30–4.60"]
null
0.034
d
Ground-truth self-score 1 (make_gt takes the yht points directly); yht FEAT_CODE=1 LE; build orientation check best=id (original orientation has the strongest signal), in the overlay the ground-truth points lie on the side of the given PA, orientation correct. Score of the quadratic fit against the ground truth itself ...
r1/r2 follow the ground-truth front in C2 and the early part of C3; r1 stops tracking at about 5.8 R☉ (coverage 69%); r3 followed a slower internal bright structure, 26% low throughout.
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screening (install scipy) and the one-look/text screening (fix file-type answers and 16-bit display).
ap-v01__ht-23badf05
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [-2160.0, -1440.0, -720.0, -360.0, 0.0, 720.0, 1800.0, 2520.0, 3240.0, 5422.0, 6121.0, 6840.0, 7560.0, 8280.0, 9000.0, 9720.0, 11160.0, 11880.0, 12600.0, 13320.0, 14040.0], "y": [2.89, 3.81, 4.62, 4.93, 5.64, 6.32, 7.63, 8.3, 8.79, 10.5, 11.94, 11.98, 12.38, 13.88, 14.73, 15.12, 16.01, 17.32, 17.48, 17.81, 18.37]...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-23badf05
[ "items/ap-v01/ht-23badf05/input/frames.csv", "items/ap-v01/ht-23badf05/input/c3/0001.png", "items/ap-v01/ht-23badf05/input/c3/0002.png", "items/ap-v01/ht-23badf05/input/c3/0003.png", "items/ap-v01/ht-23badf05/input/c3/0004.png", "items/ap-v01/ht-23badf05/input/c3/0005.png", "items/ap-v01/ht-23badf05/inp...
22,452,762
23
Expert annotation
CDAW SOHO/LASCO CME catalogue 2025/05/05 23:24:05 CPA 27 W 19 V 668.0 acc -14.9* MPA 30 | ; expert height-time file 20250505.232405.w019n.v0668.p030g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2025_05/yht/20250505.232405.w019n.v0668.p030g.yht, QUALITY_INDEX 2, observer michalek), all 21 C2/C3 points, t is ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/250506/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/250506/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2025_05/yht/20250505.232405.w019n...
items/ap-v01/ht-23badf05/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers, provide o...
After exposure normalization, do running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments and its farthest point measured, while also being distinguished from other concurrent CMEs, streamers and residual structures; one must view the entire sequence and repeatedly compare adjacent frames.
The prompt gives only the first-appearance time and MPA (or no count); uniform extrapolation at a typical speed earns only partial credit; single-value speed questions were already guessed from priors in old items, so this source uses only curves, combinations and multi-target lists. Difficulty proxy see strata.
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for item screening.
[ 0.381, 0.619, 0.429 ]
["19-point csv, t 0–14760 s, y 5.30–20.30", "15-point csv, t 0–11160 s, y 5.20–16.00", "13-point csv, t 0–11160 s, y 5.70–17.30"]
null
0.28600000000000003
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points are on the side of the given PA, direction correct. Score of the quadratic fit against the ground truth itself 1.00. Medi...
C3 ground truth is at the outer end of the bright front; r1 overshoots the front by 1–2 R☉ and overestimates, r2/r3 stop tracking after 16 R☉ (coverage 62%).
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screen (install scipy) and the one-look/text screen (fix file-type answers and 16-bit display).
ap-v01__ht-24958901
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [726.0, 2160.0, 2880.0, 4320.0, 7233.0, 8640.0, 10800.0], "y": [3.57, 4.26, 4.71, 5.79, 7.37, 8.52, 9.79]}
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-24958901
[ "items/ap-v01/ht-24958901/input/frames.csv", "items/ap-v01/ht-24958901/input/c2/0001.png", "items/ap-v01/ht-24958901/input/c2/0002.png", "items/ap-v01/ht-24958901/input/c2/0003.png", "items/ap-v01/ht-24958901/input/c2/0004.png", "items/ap-v01/ht-24958901/input/c3/0001.png", "items/ap-v01/ht-24958901/inp...
10,271,267
11
Expert annotation
CDAW SOHO/LASCO CME catalogue 2002/12/01 03:30:11 CPA 80 W 20 V 437.0 acc 5.9* MPA 84 | ; expert height-time file 20021201.033011.w020n.v0437.p084g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2002_12/yht/20021201.033011.w020n.v0437.p084g.yht, QUALITY_INDEX 2, observer michalek), all 7 C2/C3 points, t is sec...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/021201/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/021201/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2002_12/yht/20021201.033011.w020n...
items/ap-v01/ht-24958901/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard the FITS header, provide...
After exposure normalization, do running difference; frame by frame, find the outer edge of the front along a radial profile near the given PA, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured; it must also be distinguished from other concurrent CMEs, streamers, and residual structures, so the whole sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or does not give the count); constant-speed extrapolation with a typical speed earns only partial credit; single-value speed tasks were already guessed from priors in old tasks, so this source uses only curves, combinations, and multi-target lists. For di...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for task screening.
[ 0.714, 0.143, 0.429 ]
["9-point csv, t 726–12240 s, y 2.50–12.00", "8-point csv, t 2160–12240 s, y 3.90–9.00", "9-point csv, t 726–12240 s, y 2.30–12.30"]
null
0.14300000000000002
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points are on the side of the given PA, direction correct. Score of the quadratic fit against the ground truth itself 1.00. Medi...
Narrow CME (W20, 7 ground-truth points): C2 ground truth at the end of the bright streak; r2 is low by about 12% in C3, r1/r3 overestimate in the final part of C3.
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screening (install scipy) and the one-look/text screening (fix file-type answers and 16-bit display).
ap-v01__ht-4263dc47
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [2161.0, 3600.0, 5040.0, 5760.0, 7222.0, 8641.0, 10895.0, 12241.0, 14400.0, 19443.0, 21600.0, 25200.0, 26640.0, 27360.0, 28800.0, 30240.0, 32494.0, 34560.0, 36000.0, 38176.0, 39604.0, 41040.0, 43439.0, 44793.0, 46080.0, 47520.0, 49680.0, 51120.0, 54721.0, 56880.0, 58321.0, 60481.0, 61920.0, 64080.0, 65521.0, 6768...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-4263dc47
[ "items/ap-v01/ht-4263dc47/input/frames.csv", "items/ap-v01/ht-4263dc47/input/c3/0001.png", "items/ap-v01/ht-4263dc47/input/c3/0002.png", "items/ap-v01/ht-4263dc47/input/c3/0003.png", "items/ap-v01/ht-4263dc47/input/c3/0004.png", "items/ap-v01/ht-4263dc47/input/c3/0005.png", "items/ap-v01/ht-4263dc47/inp...
72,075,359
75
Expert annotation
CDAW SOHO/LASCO CME catalogue 2009/01/30 23:06:04 CPA 308 W 7 V 102.0 acc 2.7* MPA 282 | ; expert height-time file 20090130.230604.w007n.v0102.p282g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2009_01/yht/20090130.230604.w007n.v0102.p282g.yht, QUALITY_INDEX 2, observer michalek), all 56 C2/C3 points, t in s...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/090130/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/090130/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/090131/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/090131/c3/ (level-0.5 FITS)", "ht...
items/ap-v01/ht-4263dc47/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using the FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers,...
After exposure normalization, make running differences; along the radial profile near the given PA, find the outer edge of the leading front frame by frame, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also distinguishing it from other concurrent CMEs, streamers and residual structures; the whole sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or does not give the count); uniform extrapolation at a typical speed earns only partial credit; single-value speed tasks were already guessed from priors in old tasks, so this source uses only curves, combinations and multi-target lists. Difficulty proxy...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) was pre-run to screen tasks.
[ 0.446, 0.482, 0.393 ]
["54-point csv, t 2161–92880 s, y 2.60–9.67", "42-point csv, t 2161–90721 s, y 3.20–8.64", "35-point csv, t 2161–90721 s, y 2.70–7.90"]
null
0
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has strongest signal), in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of a quadratic fit to the ground truth itself: 0.80. Median deviati...
All three runs correct in C2; after entering C3 all three consistently follow an internal bright structure, reaching only 8–10 R☉ at the end (ground truth 21 R☉). In the polar sector plots the ground truth is at the outer end of the front in every frame.
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screen (install scipy) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__ht-7a7c8ea8
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 512x512 px, 112.0 arcsec/px), 16-bit PN...
file (csv: t,y)
{"t": [2875.0, 3606.0, 4316.0, 5046.0, 5755.0, 6486.0, 8739.0, 9370.0, 10075.0, 10806.0, 11517.0, 12246.0, 12955.0, 13687.0, 14396.0, 15126.0, 15835.0, 16566.0, 17277.0, 18006.0, 18717.0, 19446.0, 20155.0, 20887.0, 21596.0, 22326.0, 23035.0, 23766.0, 24475.0, 26649.0, 27356.0, 28086.0, 28797.0, 29527.0, 30236.0, 30967....
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-7a7c8ea8
[ "items/ap-v01/ht-7a7c8ea8/input/frames.csv", "items/ap-v01/ht-7a7c8ea8/input/c3/0001.png", "items/ap-v01/ht-7a7c8ea8/input/c3/0002.png", "items/ap-v01/ht-7a7c8ea8/input/c3/0003.png", "items/ap-v01/ht-7a7c8ea8/input/c3/0004.png", "items/ap-v01/ht-7a7c8ea8/input/c3/0005.png", "items/ap-v01/ht-7a7c8ea8/inp...
76,006,494
147
Expert annotation
CDAW SOHO/LASCO CME catalogue 2026/07/21 23:48:20 CPA 202 W 146 V 238.0 acc 10.8 MPA 202 | Partial Halo; expert height-time file 20260721.234820.w146p.v0238.p202g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2026_07/yht/20260721.234820.w146p.v0238.p202g.yht, QUALITY_INDEX 4, observer michalek), all 99 C2/C3 ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/260721/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/260721/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/260722/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/260722/c3/ (level-0.5 FITS)", "ht...
items/ap-v01/ht-7a7c8ea8/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024 frames with Clear polarization (C2 Orange, C3 Clear filter); using the FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun, and resample to the nominal pixel scale (C2 1024², C3 512²), 16-bit PNG, FITS headers discard...
After exposure normalization, do running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also distinguishing it from other concurrent CMEs, streamers, and residual structures; this requires viewing the entire sequence and repeatedly comparing adjacent frames.
The prompt gives only the first-appearance time and MPA (or gives no number); constant-speed extrapolation at a typical speed earns only partial credit; single-value speed questions were already guessed from priors in old questions, so this source uses only curves, combinations, and multi-target lists. For the difficul...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is used for pre-screening questions.
[ 0.394, 0.778, 0.162 ]
["78-point csv, t 10806–68022 s, y 3.04–15.75", "91-point csv, t 2875–74519 s, y 2.77–28.40", "94-point csv, t 2875–74519 s, y 2.30–27.95"]
null
0
d
Ground-truth self-score 1 (make_gt takes the yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal); in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of a quadratic fit to the ground truth itself: 0.69. Median...
r2 close (0.78); r1 stops tracking at C3 15 R☉, r3 follows internal structure in C3, 16% low.
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screen (install scipy) and the one-look/text screen (fix file-type answers and 16-bit display).
ap-v01__ht-84dbadec
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 4320.0, 5042.0, 5400.0, 5760.0, 6120.0, 6841.0, 7560.0, 8280.0, 9000.0, 9720.0, 10442.0, 11160.0, 11880.0, 12600.0, 13320.0, 14040.0, 14760.0, 15481.0, 16200.0, 16920.0, 17640.0, 18360.0, 19080.0], "y": [2.75, 4.0, 5.15, 5.84, 6.32, 7.1, 8.35, 9.44, 10.59, 11.24, 12.66, 13.61, 14.67, 15.46, 16.32, 17.6, ...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-84dbadec
[ "items/ap-v01/ht-84dbadec/input/frames.csv", "items/ap-v01/ht-84dbadec/input/c3/0001.png", "items/ap-v01/ht-84dbadec/input/c3/0002.png", "items/ap-v01/ht-84dbadec/input/c3/0003.png", "items/ap-v01/ht-84dbadec/input/c3/0004.png", "items/ap-v01/ht-84dbadec/input/c3/0005.png", "items/ap-v01/ht-84dbadec/inp...
36,530,446
37
Expert annotation
CDAW SOHO/LASCO CME catalogue 2013/12/12 03:36:05 CPA 214 W 276 V 1002.0 acc -15.3 MPA 236 | Partial Halo; expert height-time file 20131212.033605.w276p.v1002.p236g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2013_12/yht/20131212.033605.w276p.v1002.p236g.yht, QUALITY_INDEX 5, observer michalek), all 24 C2/C...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/131212/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/131212/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2013_12/yht/20131212.033605.w276p...
items/ap-v01/ht-84dbadec/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun, and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers, provide ...
After exposure normalization, compute running difference; frame by frame, find the outer edge of the front along a radial profile near the given PA, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also being distinguished from other concurrent CMEs, streamers, and residual structures; the whole sequence must be viewed with repeated comparison between adjacent frames.
The prompt gives only the first-appearance time and MPA (or does not give the count); constant-speed extrapolation at a typical speed earns only partial credit; single-value speed questions were already guessed from priors in old tasks, so this source uses only curves, combinations, and multi-target lists. For the diff...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for task screening.
[ 0.25, 0.375, 0.333 ]
["14-point csv, t 3600–11160 s, y 2.55–12.60", "12-point csv, t 3600–9720 s, y 2.50–11.90", "19-point csv, t 3600–14040 s, y 2.50–17.60"]
null
0.042
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of a quadratic fit to the ground truth itself: 1.00. Median dev...
Ground truth is at the faint outer edge beyond the bright loop; all three stop tracking at about 13 R☉ (coverage 46–71%), C2 segment low by 8–9%.
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screen (install scipy) and the one-look/text screen (fix file-type answers and 16-bit display).
ap-v01__ht-8df36389
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 4320.0, 5040.0, 5760.0, 8005.0, 8280.0, 9000.0, 9720.0, 10440.0, 11160.0, 11881.0, 12600.0, 14042.0, 14761.0, 15480.0, 16200.0, 16920.0, 17640.0], "y": [2.54, 3.13, 3.74, 4.22, 5.97, 6.08, 6.97, 7.4, 7.81, 7.79, 8.97, 8.97, 9.71, 10.88, 11.83, 11.86, 12.84, 13.58]}
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-8df36389
[ "items/ap-v01/ht-8df36389/input/frames.csv", "items/ap-v01/ht-8df36389/input/c3/0001.png", "items/ap-v01/ht-8df36389/input/c3/0002.png", "items/ap-v01/ht-8df36389/input/c3/0003.png", "items/ap-v01/ht-8df36389/input/c3/0004.png", "items/ap-v01/ht-8df36389/input/c3/0005.png", "items/ap-v01/ht-8df36389/inp...
33,962,251
34
Expert annotation
CDAW SOHO/LASCO CME catalogue 2023/12/20 00:12:05 CPA 249 W 156 V 524.0 acc 6.1* MPA 266 | Fixed the MVI roll angle issue on 2026-01-30; Partial Halo; expert height-time file 20231220.001205.w156p.v0524.p266g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2023_12/yht/20231220.001205.w156p.v0524.p266g.yht, QUAL...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/231219/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/231219/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/231220/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/231220/c3/ (level-0.5 FITS)", "ht...
items/ap-v01/ht-8df36389/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear polarization (C2 Orange, C3 Clear filter) frames; using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard the FITS header, provide on...
After exposure normalization, do running difference; near the PA given in the question, find the outer edge of the front along a radial profile frame by frame, convert to R☉, join C2 and C3 into a height–time curve, exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured; it must also be distinguished from other concurrent CMEs, streamers and residual structures. The entire sequence must be viewed, with repeated comparison between adjacent frames.
The question gives only the first-appearance time and MPA (or no number); constant-speed extrapolation at a typical speed earns only partial credit; single-value speed questions were already guessed from priors in old questions, so this source uses only curve, combination and multi-target list questions. Difficulty pro...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for question screening.
[ 0.278, 0.5, 0.556 ]
["20-point csv, t 3600–19080 s, y 2.60–8.50", "21-point csv, t 3600–19080 s, y 2.40–10.80", "14-point csv, t 3600–14042 s, y 2.45–9.60"]
null
0
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points are on the side of the question's PA, direction correct. Score of the quadratic fit against the ground truth itself 1.00....
C3 ground truth is at the diffuse outer boundary, the three runs follow the bright core, C3 low by 6–12% (more for r1).
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screening (install scipy) and the one-look/text screening (fix file-type answers and 16-bit display).
ap-v01__ht-93499592
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [2882.0, 3600.0, 4320.0, 5040.0, 6483.0, 6840.0, 7201.0, 7561.0, 7920.0, 8280.0, 9002.0, 9720.0, 10441.0, 11160.0, 11880.0, 12600.0, 13322.0, 14040.0, 14762.0, 15480.0, 16201.0, 16920.0, 17640.0, 18360.0, 19080.0], "y": [2.85, 3.21, 3.86, 4.23, 5.33, 5.3, 5.77, 5.79, 6.25, 6.2, 6.68, 7.35, 7.7, 7.89, 8.38, 8.78, ...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-93499592
[ "items/ap-v01/ht-93499592/input/frames.csv", "items/ap-v01/ht-93499592/input/c2/0001.png", "items/ap-v01/ht-93499592/input/c2/0002.png", "items/ap-v01/ht-93499592/input/c2/0003.png", "items/ap-v01/ht-93499592/input/c2/0004.png", "items/ap-v01/ht-93499592/input/c2/0005.png", "items/ap-v01/ht-93499592/inp...
38,509,417
39
Expert annotation
CDAW SOHO/LASCO CME catalogue 2015/08/08 08:12:06 CPA 36 W 18 V 432.0 acc 5.0* MPA 31 | ; expert height-time file 20150808.081206.w018n.v0432.p031g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2015_08/yht/20150808.081206.w018n.v0432.p031g.yht, QUALITY_INDEX 2, observer gmichale), all 25 C2/C3 points, t is se...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/150808/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/150808/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2015_08/yht/20150808.081206.w018n...
items/ap-v01/ht-93499592/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard the FITS headers, pr...
After exposure normalization, do running difference; near the given PA, find the outer edge of the leading front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured; it must also be distinguished from other concurrent CMEs, streamers and residual structures, which requires viewing the entire sequence and repeatedly comparing adjacent frames.
The prompt gives only the first-appearance time and MPA (or does not give the count); constant-velocity extrapolation at a typical speed can earn only partial credit; single-value speed questions were already guessed from priors in old questions, so this source uses only curves, combinations, and multi-target lists. Fo...
CORIMP (IDL) and CACTus/SEEDS cannot be run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for question screening.
[ 0.04, 0.08, 0.36 ]
["17-point csv, t 2882–14040 s, y 3.26–12.60", "18-point csv, t 2882–14762 s, y 3.00–12.50", "15-point csv, t 3600–12600 s, y 3.30–10.20"]
null
0.12
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has strongest signal), in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of the quadratic fit against the ground truth itself 1.00. Median d...
C3 ground truth is at the visible tip of the jet; r1/r2 overshoot the tip in C3, overestimating by 18–30%; r3 covers only 60%.
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screen (install scipy) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__ht-9a5fb8b8
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 4322.0, 5568.0, 6485.0, 6840.0, 7200.0, 7560.0, 7920.0, 8280.0, 9001.0, 9720.0, 10440.0, 11160.0, 14040.0, 14760.0, 15480.0, 16200.0, 18361.0, 19080.0, 19800.0, 21983.0, 22680.0, 23401.0, 24120.0, 24840.0, 25561.0, 26280.0], "y": [2.63, 3.0, 4.12, 4.96, 5.38, 5.88, 6.34, 6.42, 7.12, 7.44, 7.89, 8.76, 9.6...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-9a5fb8b8
[ "items/ap-v01/ht-9a5fb8b8/input/frames.csv", "items/ap-v01/ht-9a5fb8b8/input/c2/0001.png", "items/ap-v01/ht-9a5fb8b8/input/c2/0002.png", "items/ap-v01/ht-9a5fb8b8/input/c2/0003.png", "items/ap-v01/ht-9a5fb8b8/input/c2/0004.png", "items/ap-v01/ht-9a5fb8b8/input/c2/0005.png", "items/ap-v01/ht-9a5fb8b8/inp...
40,881,248
41
Expert annotation
CDAW SOHO/LASCO CME catalogue 2023/04/16 20:24:05 CPA 121 W 118 V 782.0 acc 17.8 MPA 123 | Fixed the MVI roll angle issue on 2026-01-30; expert height-time file 20230416.202405.w118n.v0782.p123g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2023_04/yht/20230416.202405.w118n.v0782.p123g.yht, QUALITY_INDEX 3, o...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/230416/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/230416/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/230417/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https...
items/ap-v01/ht-9a5fb8b8/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers, provide o...
After exposure normalization, do running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, stitch C2 and C3 into a height–time curve, exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments and its farthest point measured, and it must also be distinguished from other concurrent CMEs, streamers and residual structures; the whole sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or does not give the count); extrapolating at constant typical speed earns only partial credit; single-value speed questions were already guessed from priors in old tasks, so this source uses only curves, combinations and multi-target lists. Difficulty pr...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) was pre-run for screening tasks.
[ 0.63, 0.259, 0.111 ]
["24-point csv, t 3600–23401 s, y 2.47–23.60", "18-point csv, t 3600–21983 s, y 2.60–19.50", "29-point csv, t 3600–27720 s, y 2.50–22.80"]
null
0.037
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points are on the side of the given PA, direction correct. Score of the quadratic-fit ground truth against itself 1.00. Median d...
All three low in C2; r2/r3 low by 11–17% in C3; r1's C3 segment is close.
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screen (install scipy) and the one-look/text screen (fix file-type answers and 16-bit display).
ap-v01__ht-ad4d9ecd
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [2981.0, 3610.0, 4350.0, 5040.0, 5760.0, 6481.0, 7200.0, 7560.0, 7920.0, 9000.0, 9363.0, 9721.0, 10080.0, 10440.0, 11161.0, 11881.0, 12601.0, 13320.0, 14041.0, 14761.0, 15481.0, 16201.0, 16920.0, 17640.0, 18360.0, 19080.0, 19803.0, 20522.0, 21240.0, 21961.0, 22680.0, 23400.0], "y": [3.03, 3.4, 3.8, 4.11, 4.32, 4....
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-ad4d9ecd
[ "items/ap-v01/ht-ad4d9ecd/input/frames.csv", "items/ap-v01/ht-ad4d9ecd/input/c3/0001.png", "items/ap-v01/ht-ad4d9ecd/input/c3/0002.png", "items/ap-v01/ht-ad4d9ecd/input/c3/0003.png", "items/ap-v01/ht-ad4d9ecd/input/c3/0004.png", "items/ap-v01/ht-ad4d9ecd/input/c3/0005.png", "items/ap-v01/ht-ad4d9ecd/inp...
43,152,939
44
Expert annotation
CDAW SOHO/LASCO CME catalogue 2015/05/03 01:25:46 CPA 245 W 75 V 273.0 acc 4.9* MPA 244 | ; expert height-time file 20150503.012546.w075n.v0273.p244g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2015_05/yht/20150503.012546.w075n.v0273.p244g.yht, QUALITY_INDEX 2, observer michalek), all 32 C2/C3 points, t is ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/150503/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/150503/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2015_05/yht/20150503.012546.w075n...
items/ap-v01/ht-ad4d9ecd/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear polarization (C2 Orange, C3 Clear filter) frames; using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers, provide on...
After exposure normalization, do running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, stitch C2 and C3 into a height–time curve, exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also being distinguished from other concurrent CMEs, streamers and residual structures; the whole sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or does not give the count); uniform extrapolation at a typical speed earns only partial credit; single-value speed tasks were already guessed from priors in old tasks, so this source uses only curves, combinations, and multi-target lists. Difficulty prox...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for task screening.
[ 0.281, 0.062, 0.875 ]
["12-point csv, t 2981–11521 s, y 3.00–5.90", "31-point csv, t 2981–24840 s, y 3.00–10.60", "34-point csv, t 2981–24840 s, y 2.85–12.29"]
null
0.062
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original orientation has strongest signal), in the overlay the ground-truth points are on the side of the given PA, orientation correct. Score of the quadratic fit against the ground truth itself 0.97. Medi...
C2 ground truth is on the faint front outside the bright streamer: r3 close (0.875); r1 stops tracking at 6 R☉ (coverage 47%); r2 is 14% low throughout.
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screen (install scipy) and the one-look/text screen (fix file-type answers and 16-bit display).
ap-v01__ht-b11a81cd
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 5045.0, 5760.0, 6481.0, 7200.0, 7920.0, 8640.0, 9361.0, 9720.0, 10080.0, 10440.0, 10800.0, 11160.0, 11520.0, 11881.0, 12600.0, 13321.0, 14040.0, 14760.0, 15481.0, 16200.0, 16921.0, 17640.0, 18361.0, 19080.0, 19800.0, 20520.0, 21240.0, 21962.0, 22680.0, 23400.0, 24121.0, 24840.0, 26281.0, 27000.0, 27720.0...
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-b11a81cd
[ "items/ap-v01/ht-b11a81cd/input/frames.csv", "items/ap-v01/ht-b11a81cd/input/c2/0001.png", "items/ap-v01/ht-b11a81cd/input/c2/0002.png", "items/ap-v01/ht-b11a81cd/input/c2/0003.png", "items/ap-v01/ht-b11a81cd/input/c2/0004.png", "items/ap-v01/ht-b11a81cd/input/c2/0005.png", "items/ap-v01/ht-b11a81cd/inp...
57,198,452
58
Expert annotation
CDAW SOHO/LASCO CME catalogue 2013/08/06 02:48:05 CPA 188 W 99 V 632.0 acc 25.5 MPA 188 | ; expert height-time file 20130806.024805.w099n.v0632.p188g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2013_08/yht/20130806.024805.w099n.v0632.p188g.yht, QUALITY_INDEX 5, observer michalek), all 39 C2/C3 points, t is ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/130806/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/130806/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2013_08/yht/20130806.024805.w099n...
items/ap-v01/ht-b11a81cd/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keep only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers, provide on...
After exposure normalization, do running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, exclude other simultaneous CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also distinguishing it from other simultaneous CMEs, streamers and residual structures; the whole sequence must be viewed with repeated comparison between adjacent frames.
The task statement gives only the first-appearance time and MPA (or not the count); constant-velocity extrapolation at a typical speed earns only partial credit; single-value speed questions were already guessed from priors in older tasks, so this source uses only curves, combinations and multi-target lists. Difficulty...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for task screening.
[ 0.205, 0.462, 0.513 ]
["39-point csv, t 3600–29160 s, y 2.05–15.41", "25-point csv, t 5045–26281 s, y 2.55–13.60", "25-point csv, t 5045–19080 s, y 2.64–12.93"]
null
0.026000000000000002
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal), in the overlay the ground-truth points are on the side of the given PA, direction correct. Score of a quadratic fit to the ground truth itself: 0.87. Median dev...
r3 tracks to 12 R☉ then stops tracking; r1/r2 follow an internal bright structure in C3, too low. The late C3 front (>16 R☉) is not evident in this audit rendering, but QUALITY_INDEX 5, and r3 independently reproduces it up to 12 R☉.
Keep, evaluation-set candidate; before entering the evaluation set, must rerun the specialist-model screen (install scipy) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__ht-bd38d929
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 4320.0, 5040.0, 5400.0, 5760.0, 6120.0, 6840.0, 7560.0, 8280.0, 9000.0, 9720.0, 10440.0, 11880.0], "y": [2.52, 3.12, 4.2, 5.12, 5.32, 5.91, 7.04, 8.06, 8.87, 9.77, 11.25, 12.19, 13.44]}
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-bd38d929
[ "items/ap-v01/ht-bd38d929/input/frames.csv", "items/ap-v01/ht-bd38d929/input/c2/0001.png", "items/ap-v01/ht-bd38d929/input/c2/0002.png", "items/ap-v01/ht-bd38d929/input/c2/0003.png", "items/ap-v01/ht-bd38d929/input/c2/0004.png", "items/ap-v01/ht-bd38d929/input/c2/0005.png", "items/ap-v01/ht-bd38d929/inp...
27,474,936
27
Expert annotation
CDAW SOHO/LASCO CME catalogue 2023/10/19 06:48:05 CPA 73 W 19 V 966.0 acc -23.2 MPA 77 | Fixed the MVI roll angle issue on 2026-01-30; expert height-time file 20231019.064805.w019n.v0966.p077g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2023_10/yht/20231019.064805.w019n.v0966.p077g.yht, QUALITY_INDEX 3, obs...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/231019/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/231019/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2023_10/yht/20231019.064805.w019n...
items/ap-v01/ht-bd38d929/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024 frames with Clear polarization (C2 Orange, C3 Clear filter); using the FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard the FITS ...
After exposure normalization, take running differences; along radial profiles near the given PA, find the outer edge of the front in each frame, convert to R☉, join C2 and C3 into one height-time curve, and exclude other CMEs at the same time.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and measured at its farthest point; it must also be distinguished from other concurrent CMEs, streamers and residual structures, which requires viewing the entire sequence and repeatedly comparing adjacent frames.
The question gives only the first-appearance time and MPA (or not the number of events); extrapolating at a typical constant speed earns only partial credit; single-value speed questions were already guessed from priors in the earlier item set, so this source uses only curves, combined quantities and multi-target lists...
CORIMP (IDL) and CACTus/SEEDS cannot be run directly on the cluster; a self-written running-difference front tracker/detector with a fixed configuration (specialist.py) was run beforehand to screen items.
[ 0.538, 0.538, 0.385 ]
["10-point csv, t 4320–9720 s, y 3.05–9.50", "8-point csv, t 3600–7560 s, y 2.60–8.20", "11-point csv, t 3600–9720 s, y 2.70–11.50"]
null
0.308
d
Ground-truth self-score 1 (make_gt takes the yht points directly); yht FEAT_CODE=1 LE; the build-time direction check gives best=id (strongest signal in the original direction), and in the overlay the ground-truth points lie on the side of the given PA, so the direction is correct. Score of a quadratic fit to the groun...
All three runs are close to the ground truth, but tracking stops after about 11 R☉ (coverage 62–85%); r1 is low in the final segment.
Keep as benchmark candidate; before entering the benchmark, rerun the specialist-model screen (with scipy installed) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__ht-ddda4aac
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3588.0, 4318.0, 4670.0, 5374.0, 6110.0, 6814.0, 7550.0, 8254.0, 8990.0, 9694.0, 10430.0, 11134.0, 11870.0, 12574.0, 13310.0, 14015.0, 14750.0], "y": [3.87, 6.02, 6.18, 8.45, 9.75, 11.55, 13.06, 14.94, 15.94, 17.83, 19.48, 20.44, 21.83, 23.63, 24.54, 26.75, 28.08]}
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-ddda4aac
[ "items/ap-v01/ht-ddda4aac/input/frames.csv", "items/ap-v01/ht-ddda4aac/input/c3/0001.png", "items/ap-v01/ht-ddda4aac/input/c3/0002.png", "items/ap-v01/ht-ddda4aac/input/c3/0003.png", "items/ap-v01/ht-ddda4aac/input/c3/0004.png", "items/ap-v01/ht-ddda4aac/input/c3/0005.png", "items/ap-v01/ht-ddda4aac/inp...
26,448,306
27
Expert annotation
CDAW SOHO/LASCO CME catalogue 2026/05/10 13:48:21 CPA Halo W 360 V 1486.0 acc -27.4 MPA 40 | ; expert height-time file 20260510.134821.w360h.v1486.p040g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2026_05/yht/20260510.134821.w360h.v1486.p040g.yht, QUALITY_INDEX 5, observer michalek), all 17 C2/C3 points, t ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/260510/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/260510/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2026_05/yht/20260510.134821.w360h...
items/ap-v01/ht-ddda4aac/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024, Clear-polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to solar north up / east left, center on Sun center and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard FITS headers,...
After exposure normalization, compute running differences, find the outer edge of the front frame by frame along a radial profile near the given PA, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while also being distinguished from other concurrent CMEs, streamers, and residual structures; the entire sequence must be viewed and adjacent frames compared repeatedly.
The task statement gives only the first-appearance time and MPA (or not the count); constant-velocity extrapolation at a typical speed earns only partial credit; single-value speed tasks were already guessed from priors in earlier tasks, so this source uses only curves, combinations, and multi-target lists. Difficulty ...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is used for pre-run screening.
[ 0.471, 0.412, 0.471 ]
["9-point csv, t 3588–8990 s, y 3.63–15.05", "7-point csv, t 3588–7550 s, y 3.80–14.00", "10-point csv, t 3588–9694 s, y 3.58–15.80"]
null
0
d
Ground-truth self-score 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has strongest signal), in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of the quadratic-fit ground truth against itself 1.00. Median devia...
All three runs agree with ground truth over the covered segment, but all stop tracking at 13–16 R☉, while ground truth extends to 28 R☉ (coverage 41–59%).
Keep, evaluation-set candidate; before entering the evaluation set, must re-run specialist-model screening (install scipy) and one-look/text screening (fix file-type answers and 16-bit display).
ap-v01__ht-df3b9283
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [2830.0, 3777.0, 4719.0, 5672.0, 6616.0, 7567.0, 8509.0, 9453.0, 10391.0, 11335.0, 12282.0, 13225.0, 14166.0, 15106.0, 16056.0, 16996.0, 17940.0, 18877.0, 19829.0, 20766.0, 21707.0, 22647.0, 23600.0, 24537.0, 25477.0, 26416.0, 26679.0, 27359.0, 27619.0, 28283.0, 28546.0, 29225.0, 29490.0, 30164.0, 30426.0, 31104....
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-df3b9283
[ "items/ap-v01/ht-df3b9283/input/frames.csv", "items/ap-v01/ht-df3b9283/input/c2/0001.png", "items/ap-v01/ht-df3b9283/input/c2/0002.png", "items/ap-v01/ht-df3b9283/input/c2/0003.png", "items/ap-v01/ht-df3b9283/input/c2/0004.png", "items/ap-v01/ht-df3b9283/input/c2/0005.png", "items/ap-v01/ht-df3b9283/inp...
108,825,275
121
Expert annotation
CDAW SOHO/LASCO CME catalogue 1997/11/08 21:36:49 CPA 263 W 72 V 185.0 acc 4.0* MPA 254 | Newly inserted on 2024/05/01.; expert height-time file 19971108.213649.w072n.v0185.p254g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/1997_11/yht/19971108.213649.w072n.v0185.p254g.yht, QUALITY_INDEX 2, observer michalek...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/971108/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/971108/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/971109/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/971109/c3/ (level-0.5 FITS)", "ht...
items/ap-v01/ht-df3b9283/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024 frames with Clear polarization (C2 Orange, C3 Clear filters); using the FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun's center and resample to the nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, discard th...
After exposure normalization, take running differences; along radial profiles near the given PA, find the outer edge of the front in each frame, convert to R☉, join C2 and C3 into one height-time curve, and exclude other CMEs at the same time.
The leading front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and measured at its farthest point; it must also be distinguished from other concurrent CMEs, streamers and residual structures, so the entire sequence must be viewed and adjacent frames repeatedly compared.
The question gives only the first-appearance time and MPA (or not the number of events); extrapolating at a typical constant speed earns only partial credit; single-value speed questions were already guessed from priors in the earlier item set, so this source uses only curves, combined quantities and multi-target lists...
CORIMP (IDL) and CACTus/SEEDS cannot be run directly on the cluster; a self-written running-difference front tracker/detector with a fixed configuration (specialist.py) was run beforehand to screen items.
[ 0.409, 0.17, 0.159 ]
["50-point csv, t 11335–56731 s, y 2.64–11.54", "73-point csv, t 3777–85521 s, y 2.62–15.12", "69-point csv, t 2830–86298 s, y 2.40–17.16"]
null
0.011
d
Ground-truth self-score 1 (make_gt takes the yht points directly); yht FEAT_CODE=1 LE; the build-time direction check gives best=id (strongest signal in the original direction), and in the overlay the ground-truth points lie on the side of the given PA, so the direction is correct. Score of a quadratic fit to the groun...
r1 follows the ground truth up to 12 R☉ and then stops tracking; r2/r3 follow an inner bright loop in C3 and are 21-38% low. The outer edge in the later C3 frames is faint in the rendering; r1 independently reproduces it up to 12 R☉.
Keep as benchmark candidate; before entering the benchmark, rerun the specialist-model screen (with scipy installed) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__ht-e12c79a9
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
ht_curve
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
file (csv: t,y)
{"t": [3600.0, 4322.0, 5040.0, 6480.0, 7200.0, 7920.0, 8642.0, 9360.0, 10080.0, 10800.0, 11520.0, 13776.0, 14408.0, 15120.0, 15841.0, 16560.0, 17280.0, 18002.0, 18720.0, 20166.0, 20880.0, 21601.0, 22320.0, 23040.0, 23760.0, 24481.0, 25200.0, 25922.0, 26640.0, 27360.0, 28080.0, 28800.0, 29520.0, 29880.0, 30240.0, 30600....
solar radii vs seconds
{"type": "timeseries", "value": 0.08, "mode": "rel", "floor": 0.25}
items/ap-v01/ht-e12c79a9
[ "items/ap-v01/ht-e12c79a9/input/frames.csv", "items/ap-v01/ht-e12c79a9/input/c2/0001.png", "items/ap-v01/ht-e12c79a9/input/c2/0002.png", "items/ap-v01/ht-e12c79a9/input/c2/0003.png", "items/ap-v01/ht-e12c79a9/input/c2/0004.png", "items/ap-v01/ht-e12c79a9/input/c2/0005.png", "items/ap-v01/ht-e12c79a9/inp...
121,976,086
123
Expert annotation
CDAW SOHO/LASCO CME catalogue 2020/07/05 10:36:05 CPA 90 W 90 V 209.0 acc 8.1 MPA 102 | Fixed the MVI roll angle issue on 2026-01-30; expert height-time file 20200705.103605.w090n.v0209.p102g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2020_07/yht/20200705.103605.w090n.v0209.p102g.yht, QUALITY_INDEX 3, obse...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/200705/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/200705/c3/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/200706/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https...
items/ap-v01/ht-e12c79a9/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only 1024×1024, Clear polarization frames (C2 Orange, C3 Clear filter); using FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on Sun center and resample to nominal pixel scale (C2 1024², C3 1024²), 16-bit PNG, FITS headers discarded, only ...
After exposure normalization, compute running difference; near the given PA, find the outer edge of the front frame by frame along a radial profile, convert to R☉, join C2 and C3 into a height–time curve, and exclude other concurrent CMEs.
The front must be identified frame by frame across dozens of C2→C3 frames, linked across instruments, and its farthest point measured, while being distinguished from other concurrent CMEs, streamers and residual structures; the entire sequence must be viewed with repeated comparison between adjacent frames.
The prompt gives only the first-appearance time and MPA (or does not give the count); extrapolating at constant typical speed earns only partial credit; single-value speed questions were already guessed from priors in old tasks, so this source uses only curves, combinations, and multi-target lists. For difficulty proxy...
CORIMP (IDL) and CACTus/SEEDS cannot run directly on the cluster; a self-written fixed-configuration running-difference front tracker/detector (specialist.py) is pre-run for screening.
[ 0.346, 0.179, 0.308 ]
["87-point csv, t 3600–66601 s, y 2.60–17.60", "64-point csv, t 3600–48608 s, y 2.51–10.89", "41-point csv, t 15120–58681 s, y 2.52–16.22"]
null
0.013000000000000001
d
Ground-truth self-rating 1 (make_gt takes yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has strongest signal), in the overlay the ground-truth points lie on the side of the given PA, direction correct. Score of the quadratic fit against the ground truth itself: 0.95. Median...
Ground truth is at the outer edge of the bright front and the faint tip; all three runs are systematically low by about 9–10%, just outside the 8% tolerance.
Keep, evaluation-set candidate; before entering the evaluation set, must re-run the specialist-model screen (install scipy) and one-look/text screens (fix file-type answers and 16-bit display).
ap-v01__kin-2c28adb5
ap-v01
SOHO/LASCO C2/C3 coronal mass ejections
Solar physics
kinematics
input/c2/ and input/c3/ contain raw (unprocessed, level-0.5) white-light coronagraph images from two nested coronagraphs on a spacecraft near the Sun-Earth L1 point: C2 (inner field, about 2-6 solar radii; 1024x1024 px, 11.9 arcsec/px) and C3 (outer field, about 4-30 solar radii; 1024x1024 px, 56.0 arcsec/px), 16-bit P...
object
{"speed_kms": 983, "accel_ms2": 12.0, "t_20rsun_min": 250.6}
{"speed_kms": "km/s", "accel_ms2": "m/s^2", "t_20rsun_min": "min"}
{"speed_kms": {"type": "rel", "value": 0.1}, "accel_ms2": {"type": "abs", "value": 10}, "t_20rsun_min": {"type": "abs", "value": 15}}
items/ap-v01/kin-2c28adb5
[ "items/ap-v01/kin-2c28adb5/input/frames.csv", "items/ap-v01/kin-2c28adb5/input/c3/0001.png", "items/ap-v01/kin-2c28adb5/input/c3/0002.png", "items/ap-v01/kin-2c28adb5/input/c3/0003.png", "items/ap-v01/kin-2c28adb5/input/c3/0004.png", "items/ap-v01/kin-2c28adb5/input/c3/0005.png", "items/ap-v01/kin-2c28a...
33,165,657
33
Expert annotation
CDAW SOHO/LASCO CME catalogue 2024/02/08 01:25:52 CPA 216 W 37 V 983.0 acc 12.1 MPA 214 | ; expert height-time file 20240208.012552.w037n.v0983.p214g.yht (https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2024_02/yht/20240208.012552.w037n.v0983.p214g.yht, QUALITY_INDEX 3, observer michalek), all 25 C2/C3 points, t is ...
["https://umbra.nascom.nasa.gov/pub/lasco_level05/240208/c2/ (level-0.5 FITS)", "https://umbra.nascom.nasa.gov/pub/lasco_level05/240208/c3/ (level-0.5 FITS)", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/text_ver/univ_all.txt", "https://cdaw.gsfc.nasa.gov/CME_list/UNIVERSAL_ver2/2024_02/yht/20240208.012552.w037n...
items/ap-v01/kin-2c28adb5/make_gt.py
Download level-0.5 FITS from umbra.nascom.nasa.gov, keeping only frames that are 1024×1024 with Clear polarization (C2 Orange, C3 Clear filter); using the FITS header CROTA1/CRPIX/CDELT, rotate to north-up east-left, center on the Sun's center, and resample to the nominal pixel scale (C2 1024², C3 1024²); 16-bit PNG, F...
As above, measure the front height frame by frame; a linear/quadratic fit gives the speed and acceleration, and interpolation gives the time at 20 R☉.
Speed, acceleration and the time of reaching 20 R☉ all depend on the complete frame-by-frame height series (C2+C3): the front has to be measured in every frame and joined across the two instruments, looking back to confirm that the same CME is being followed.
The question gives only the first-appearance time and MPA (or not the number of events); extrapolating at a typical constant speed earns only partial credit; single-value speed questions were already guessed from priors in the earlier item set, so this source uses only curves, combined quantities and multi-target lists...
CORIMP (IDL) and CACTus/SEEDS cannot be run directly on the cluster; a self-written running-difference front tracker/detector with a fixed configuration (specialist.py) was run beforehand to screen items.
[ 0.667, 0.667, 0 ]
[{"speed_kms": 973, "accel_ms2": -31, "t_20rsun_min": 243}, {"speed_kms": 964, "accel_ms2": -17.1, "t_20rsun_min": 252}, {"speed_kms": 841, "accel_ms2": -38.6, "t_20rsun_min": 309}]
null
0
d
Ground-truth self-assessment 1 (make_gt takes the yht points directly); yht FEAT_CODE=1 LE; build direction check best=id (original direction has the strongest signal); in the overlay the ground-truth points are on the side of the given PA, direction is correct. Ground-truth quadratic-fit acceleration standard error 3....
In r1/r2 the speed and the time of reaching 20 R☉ are within tolerance; only the quadratic-fit acceleration is wrong (2.9-4.3 tolerances off). In r3 all three quantities are wrong.
Keep as benchmark candidate; before entering the benchmark, rerun the specialist-model screen (with scipy installed) and the one-look/text screens (fix file-type answers and 16-bit display).
ap-v03__one-c2-0743926c
ap-v03
SOHO/LASCO sungrazing comets (Sungrazer Project)
Solar system astronomy
traj_one
input/frames/ contains 15 consecutive raw white-light images (1024x1024 px, grayscale PNG, frames 0000-0014) from the inner coronagraph of a two-coronagraph pair (field of view about 2-6 solar radii); the Sun is hidden behind the central occulting disk. input/frames.csv lists each frame index (0-based, matching the fil...
file (csv: t,x,y)
[[4, 160.0, 53.0], [5, 163.0, 43.0], [6, 168.0, 32.0], [7, 173.0, 22.0], [8, 177.0, 12.0], [9, 182.0, 2.0]]
px
{"type": "trajectory", "value": 4.0}
items/ap-v03/one-c2-0743926c
[ "items/ap-v03/one-c2-0743926c/input/frames.csv", "items/ap-v03/one-c2-0743926c/input/frames/0000.png", "items/ap-v03/one-c2-0743926c/input/frames/0001.png", "items/ap-v03/one-c2-0743926c/input/frames/0002.png", "items/ap-v03/one-c2-0743926c/input/frames/0003.png", "items/ap-v03/one-c2-0743926c/input/frame...
5,822,832
15
Expert annotation
Per-frame HH:MM x y (1024 browse JPEG, top-left origin) from Sungrazer Project plain-text Comet Reports (C2, moderation_state=confirmed) report [['report_bz_20130503165608']]; duplicate reports of the same comet are averaged; report times are matched to frame numbers within ±90 s. τ=4.0 px: the reporter clicked positio...
["https://soho.nascom.nasa.gov/data/REPROCESSING/Completed/2013/c2/20130503/ (15 c2_1024.jpg images from 201305031448–201305031748)", "https://sungrazer.nrl.navy.mil/index.php/plain-text-reports"]
items/ap-v03/one-c2-0743926c/make_gt.py
15 LASCO C2 1024 browse JPEGs (from about 60 minutes before the comet's first reported point to about 60 minutes after the last reported point) converted to grayscale PNG, with the date text region in the lower-left corner [975:1024, 0:371] set to 0; file names changed to sequence numbers, times changed to minutes rela...
Take the sequence median as background and difference against it, remove stars (nearly stationary between frames) and cosmic rays (single-frame), search across multiple frames for a faint point source moving in a straight line at constant speed (shift-and-stack possible), then locate it by centroid frame by frame.
A faint comet has low single-frame SNR and is mixed in with star points/cosmic rays/coronal structures; it can only be recognized by its consistent straight-line motion through frame-by-frame comparison; the track requires per-frame localization and determining the frames where it enters/leaves the field of view, so th...
No position hint; a stationary point/image center scores 0, finding only the entry point ≈0.1; straight-line motion means detection gives a near-full score, detecting the wrong target gives 0. Knowing the first and last points and drawing a straight line gives a full score; the difficulty lies in detection.
Public tool: TimeDevBlocker/soho-comet-pipeline (C2 detection + linking algorithm by the author of Tycho Tracker, designed for level-0.5 FITS); screened using the pre-run results from specialist.py.
[ 0, 0, 0 ]
[[[0.0, 355.0, 569.1], [1.0, 331.7, 584.0], [2.0, 320.1, 591.4], [3.0, 308.5, 598.9], [4.0, 296.9, 606.3], [5.0, 285.2, 613.8], [6.0, 273.6, 621.2], [7.0, 262.0, 628.6], [8.0, 250.4, 636.1], [9.0, 238.8, 643.5], [10.0, 227.2, 651.0], [11.0, 215.5, 658.4], [12.0, 203.9, 665.9], [13.0, 192.3, 673.3], [14.0, 180.7, 680.8]...
0
false
d
Ground-truth self-score 1 (1 both inline and via the track.csv file); independent recomputation (not via make_gt, taking raw points from reports.json by rid and matching frames within ±90 s) agrees point by point. gt_check stack_z=10.1, stacked SNR along the ground truth this time = 3.7. Screening: distribution baselin...
All three runs linked 'low-confidence' tracks in the streamer band at the edge of the occulter disk (r=167–371 px) and did not find the very faint incoming comet at the topmost edge of the field of view.
Enters benchmark candidates. Still missing: screening by a second model family, plus re-running a genuine single-frame screen and Qwen screen (this source idled on the file answer).
ap-v03__one-c2-353bf194
ap-v03
SOHO/LASCO sungrazing comets (Sungrazer Project)
Solar system astronomy
traj_one
input/frames/ contains 13 consecutive raw white-light images (1024x1024 px, grayscale PNG, frames 0000-0012) from the inner coronagraph of a two-coronagraph pair (field of view about 2-6 solar radii); the Sun is hidden behind the central occulting disk. input/frames.csv lists each frame index (0-based, matching the fil...
file (csv: t,x,y)
[[4, 980.0, 294.0], [5, 993.0, 288.0], [6, 1005.0, 282.0], [7, 1019.0, 277.0]]
px
{"type": "trajectory", "value": 4.0}
items/ap-v03/one-c2-353bf194
[ "items/ap-v03/one-c2-353bf194/input/frames.csv", "items/ap-v03/one-c2-353bf194/input/frames/0000.png", "items/ap-v03/one-c2-353bf194/input/frames/0001.png", "items/ap-v03/one-c2-353bf194/input/frames/0002.png", "items/ap-v03/one-c2-353bf194/input/frames/0003.png", "items/ap-v03/one-c2-353bf194/input/frame...
4,827,139
13
Expert annotation
Per-frame HH:MM x y (1024 browse JPEG, top-left origin) from Sungrazer Project plain-text Comet Reports (C2, moderation_state=confirmed) report [['report_bz_20151214184241']]; duplicate reports of the same comet are averaged; report times are matched to frame numbers within ±90 s. τ=4.0 px: points are manually clicked ...
["https://soho.nascom.nasa.gov/data/REPROCESSING/Completed/2015/c2/20151214/ (13 c2_1024.jpg images from 201512140312–201512140536)", "https://sungrazer.nrl.navy.mil/index.php/plain-text-reports"]
items/ap-v03/one-c2-353bf194/make_gt.py
13 LASCO C2 1024 browse JPEGs (from about 60 minutes before the comet's first report point to about 60 minutes after the last report point) converted to grayscale PNG, with the date text region in the lower-left corner [975:1024, 0:371] set to 0; file names changed to sequence numbers, times changed to minutes relative...
Take the sequence median as background and difference against it, remove stars (nearly stationary between frames) and cosmic rays (single-frame), search across multiple frames for a faint point source moving uniformly along a straight line (shift-and-stack possible), then locate it per frame by centroid.
A faint comet has low single-frame SNR and is mixed with star points/cosmic rays/coronal structures; only by frame-to-frame comparison can it be recognized by its consistent straight-line motion. The track requires per-frame localization and determining the frames where it enters/leaves the field of view, so the whole ...
No position hint; a stationary point/image center scores 0, finding only the entry point ≈0.1; straight-line motion means detection yields a near-full score, detecting the wrong target scores 0. Knowing the first and last points, a straight line gets full score; the difficulty lies in detection.
Public tool: TimeDevBlocker/soho-comet-pipeline (C2 detection + linking algorithm by the Tycho Tracker author, targeting level-0.5 FITS); screened using results pre-run by specialist.py.
[ 0, 0, 0 ]
[[[0.0, 118.0, 899.0], [1.0, 129.0, 870.0], [2.0, 141.0, 850.0], [3.0, 147.0, 813.0], [4.0, 164.0, 784.0], [5.0, 177.0, 754.0], [6.0, 198.0, 726.0], [7.0, 225.0, 685.0], [8.0, 237.0, 653.0], [9.0, 254.0, 620.0], [10.0, 271.0, 586.0], [11.0, 288.0, 553.0], [12.0, 309.0, 512.0]], [[0.0, 147.0, 867.0], [1.0, 94.0, 860.0],...
0
false
d
Ground-truth self-score 1 (1 for both inline and track.csv file methods); independent recomputation (not via make_gt, taking raw points from reports.json by rid and matching frames within ±90 s) agrees point by point. gt_check stack_z=25.7; stacked SNR along the ground truth this time = 7.8. Screening: distribution bas...
r1 followed the streamer in the lower left, r2 a three-frame transient, r3 another bright point in the lower right; none noticed the comet moving outward at the right edge (direction of motion opposite to the usual sungrazer entry).
Enters benchmark candidates. Still needs screening by a second model family, plus re-running genuine single-frame and Qwen screens (this source idled on the file answer).
ap-v03__one-c2-86ad7fc6
ap-v03
SOHO/LASCO sungrazing comets (Sungrazer Project)
Solar system astronomy
traj_one
input/frames/ contains 16 consecutive raw white-light images (1024x1024 px, grayscale PNG, frames 0000-0015) from the inner coronagraph of a two-coronagraph pair (field of view about 2-6 solar radii); the Sun is hidden behind the central occulting disk. input/frames.csv lists each frame index (0-based, matching the fil...
file (csv: t,x,y)
[[5, 17.0, 922.0], [6, 27.0, 913.0], [7, 40.0, 903.0], [8, 52.0, 895.0], [9, 65.0, 885.0], [10, 77.0, 876.0], [11, 89.0, 867.0], [12, 102.0, 857.0]]
px
{"type": "trajectory", "value": 4.0}
items/ap-v03/one-c2-86ad7fc6
[ "items/ap-v03/one-c2-86ad7fc6/input/frames.csv", "items/ap-v03/one-c2-86ad7fc6/input/frames/0000.png", "items/ap-v03/one-c2-86ad7fc6/input/frames/0001.png", "items/ap-v03/one-c2-86ad7fc6/input/frames/0002.png", "items/ap-v03/one-c2-86ad7fc6/input/frames/0003.png", "items/ap-v03/one-c2-86ad7fc6/input/frame...
6,444,514
16
Expert annotation
Per-frame HH:MM x y (1024 browse JPEG, top-left origin) from Sungrazer Project plain-text Comet Reports (C2, moderation_state=confirmed) report [['report_mu_20170402233649']]; duplicate reports of the same comet are averaged; report times are matched to frame numbers within ±90 s. τ=4.0 px: points are manually clicked ...
["https://soho.nascom.nasa.gov/data/REPROCESSING/Completed/2017/c2/20170402/ (16 c2_1024.jpg images from 201704021024–201704021348)", "https://sungrazer.nrl.navy.mil/index.php/plain-text-reports"]
items/ap-v03/one-c2-86ad7fc6/make_gt.py
16 LASCO C2 1024 browse JPEGs (from about 60 minutes before the comet's first report point to about 60 minutes after the last report point) converted to grayscale PNG, the lower-left date text region [975:1024, 0:371] set to 0; file names changed to sequence numbers, times changed to minutes relative to the first frame...
Take the sequence median as background and difference against it, remove stars (stars are nearly stationary between frames) and cosmic rays (single-frame), search across multiple frames for a faint point source moving uniformly along a straight line (shift-and-stack possible), then locate by centroid frame by frame.
The faint comet has low single-frame SNR and is mixed with star points/cosmic rays/coronal structures; only frame-by-frame comparison can identify it by its consistent straight-line motion; the track requires per-frame localization and determining the frames of entering/leaving the field of view, so the whole sequence ...
No position hint; a stationary point/image center scores 0, finding only the entry point ≈0.1; straight-line motion means detection yields near-full score, detecting the wrong target scores 0. Knowing the first and last points, a straight line gets full score; the difficulty is in detection.
Public tool: TimeDevBlocker/soho-comet-pipeline (C2 detection + linking algorithm by the author of Tycho Tracker, designed for level-0.5 FITS); screened using results pre-run by specialist.py.
[ 0, 0, 0 ]
[[[7.0, 632.0, 343.0], [8.0, 652.0, 371.0], [9.0, 673.0, 397.0], [10.0, 693.0, 425.0], [11.0, 713.0, 455.0], [12.0, 732.0, 485.0], [13.0, 792.0, 567.0]], [[0.0, 173.0, 48.0], [1.0, 176.0, 48.0], [2.0, 178.0, 48.0], [3.0, 180.0, 48.0], [4.0, 183.0, 48.0], [5.0, 185.0, 48.0], [6.0, 188.0, 48.0], [7.0, 190.0, 48.0], [8.0,...
0
false
d
Ground-truth self-score 1 (1 both inline and via the track.csv file); independent recomputation (not via make_gt, taking raw points from reports.json by rid and matching frames within ±90 s) agrees point by point. gt_check stack_z=8.6, stacked SNR along the ground truth this run=6.9. Screening: distribution baseline em...
r1 followed a fast linked track near the occulter disk, r2 a slowly drifting object at the upper edge, r3 a 5-frame transient; all are more than 770 px from ground truth; none found the faint comet entering at the lower-left edge.
Enters benchmark candidates. Still missing a second-model-family screen, and re-running a genuine single-frame screen and Qwen screen (this source idled on the file answer).
ap-v03__one-c2-a760bab3
ap-v03
SOHO/LASCO sungrazing comets (Sungrazer Project)
Solar system astronomy
traj_one
input/frames/ contains 17 consecutive raw white-light images (1024x1024 px, grayscale PNG, frames 0000-0016) from the inner coronagraph of a two-coronagraph pair (field of view about 2-6 solar radii); the Sun is hidden behind the central occulting disk. input/frames.csv lists each frame index (0-based, matching the fil...
file (csv: t,x,y)
[[4, 135.0, 971.0], [5, 145.0, 963.0], [6, 154.0, 955.0], [7, 162.0, 947.0], [8, 171.0, 938.0], [9, 180.0, 930.0], [10, 190.0, 922.0], [11, 200.0, 914.0]]
px
{"type": "trajectory", "value": 4.0}
items/ap-v03/one-c2-a760bab3
[ "items/ap-v03/one-c2-a760bab3/input/frames.csv", "items/ap-v03/one-c2-a760bab3/input/frames/0000.png", "items/ap-v03/one-c2-a760bab3/input/frames/0001.png", "items/ap-v03/one-c2-a760bab3/input/frames/0002.png", "items/ap-v03/one-c2-a760bab3/input/frames/0003.png", "items/ap-v03/one-c2-a760bab3/input/frame...
6,572,961
17
Expert annotation
Per-frame HH:MM x y (1024 browse JPEG, top-left origin) from Sungrazer project plain-text Comet Reports (C2, moderation_state=confirmed) report [['report_mu_20130421113521']]; duplicate reports of the same comet are averaged; report times are matched to frame numbers within ±90 s. τ=4.0 px: reporters click points manua...
["https://soho.nascom.nasa.gov/data/REPROCESSING/Completed/2013/c2/20130421/ (17 c2_1024.jpg files from 201304210812–201304211136)", "https://sungrazer.nrl.navy.mil/index.php/plain-text-reports"]
items/ap-v03/one-c2-a760bab3/make_gt.py
17 LASCO C2 1024 browse JPEGs (from about 60 minutes before the comet's first reported point to about 60 minutes after the last reported point) converted to grayscale PNG; the date-text region at the bottom-left [975:1024, 0:371] set to 0; file names changed to sequence numbers, times changed to minutes relative to the...
Take the sequence median as background and difference against it, remove stars (nearly stationary between frames) and cosmic rays (single-frame), search across multiple frames for a faint point source moving linearly at constant speed (shift-and-stack possible), then locate it per frame by centroid.
A faint comet has low single-frame SNR and is mixed in with star points/cosmic rays/coronal structures; only by comparing frames can it be recognized by its consistent linear motion; the track requires per-frame localization and determining the frames of entering/leaving the field of view, so the whole sequence must be...
No position hint; a stationary point/image center scores 0, finding only the entry point ≈0.1; linear motion means detection yields near-full score, detecting the wrong target scores 0. Knowing the first and last points allows a straight line for full score; the difficulty is detection.
Public tool: TimeDevBlocker/soho-comet-pipeline (C2 detection + linking algorithm by the Tycho Tracker author, targeting level-0.5 FITS); screened using pre-run results from specialist.py.
[ 0, 0, 0 ]
[[[9.0, 97.0, 708.0], [10.0, 123.0, 690.0], [11.0, 145.0, 677.0], [12.0, 172.0, 652.0], [13.0, 203.0, 630.0], [14.0, 229.0, 613.0], [15.0, 256.0, 594.0], [16.0, 283.0, 575.0]], [[0.0, 676.0, 401.0], [1.0, 684.0, 395.0], [2.0, 691.0, 392.0], [3.0, 698.0, 388.0], [4.0, 712.0, 380.0], [5.0, 717.0, 377.0], [6.0, 726.0, 372...
0
false
d
Ground-truth self-score 1 (1 for both inline and track.csv file methods); independent recomputation (not via make_gt, taking raw points from reports.json by rid and matching frames within ±90 s) agrees point by point. gt_check stack_z=7.4; stacked SNR along the ground truth this time = 7.8. Screening: distribution base...
r1 tracked a fast bright point in the CME region, r2 a slow point in the inner field, r3 a saturated nearly stationary bright star; none found the comet entering at the bottom-left edge.
Enters benchmark candidates. Still missing a second model-family screening, and re-running proper single-frame and Qwen screens (this source idled on the file answer).
ap-v03__one-c2-b03f2b11
ap-v03
SOHO/LASCO sungrazing comets (Sungrazer Project)
Solar system astronomy
traj_one
input/frames/ contains 13 consecutive raw white-light images (1024x1024 px, grayscale PNG, frames 0000-0012) from the inner coronagraph of a two-coronagraph pair (field of view about 2-6 solar radii); the Sun is hidden behind the central occulting disk. input/frames.csv lists each frame index (0-based, matching the fil...
file (csv: t,x,y)
[[4, 24.0, 253.0], [5, 34.0, 241.0], [6, 44.0, 228.0], [7, 53.0, 216.0]]
px
{"type": "trajectory", "value": 4.0}
items/ap-v03/one-c2-b03f2b11
[ "items/ap-v03/one-c2-b03f2b11/input/frames.csv", "items/ap-v03/one-c2-b03f2b11/input/frames/0000.png", "items/ap-v03/one-c2-b03f2b11/input/frames/0001.png", "items/ap-v03/one-c2-b03f2b11/input/frames/0002.png", "items/ap-v03/one-c2-b03f2b11/input/frames/0003.png", "items/ap-v03/one-c2-b03f2b11/input/frame...
4,934,323
13
Expert annotation
Per-frame HH:MM x y (1024 browse JPEG, top-left origin) from Sungrazer project plain-text Comet Reports (C2, moderation_state=confirmed) report [['report_rk_20180511163657']]; duplicate reports of the same comet are averaged; report times are matched to frame numbers within ±90 s. τ=4.0 px: points are manually clicked ...
["https://soho.nascom.nasa.gov/data/REPROCESSING/Completed/2018/c2/20180509/ (13 c2_1024.jpg images from 201805091812–201805092048)", "https://sungrazer.nrl.navy.mil/index.php/plain-text-reports"]
items/ap-v03/one-c2-b03f2b11/make_gt.py
13 LASCO C2 1024 browse JPEGs (from about 60 minutes before the comet's first report point to about 60 minutes after the last report point) converted to grayscale PNG, with the date-text region in the lower-left corner [975:1024, 0:371] set to 0; file names changed to sequence numbers, times changed to minutes relative...
Take the sequence median as background and difference against it, remove stars (nearly stationary between frames) and cosmic rays (single-frame), search across frames for a faint point source moving uniformly along a straight line (shift-and-stack possible), then locate it per frame by centroid.
A faint comet has low single-frame SNR and is mixed in with stars/cosmic rays/coronal structures; it can only be recognized by its consistent linear motion through frame-by-frame comparison. The track requires per-frame localization and determining the frames where it enters/leaves the field of view, so the whole seque...
No position hint; a stationary point/image center scores 0, finding only the entry point ≈0.1; linear motion means detection gives near-full score, detecting the wrong target gives 0. Knowing the first and last points allows a full score with a straight line; the difficulty is detection.
Public tool: TimeDevBlocker/soho-comet-pipeline (C2 detection + linking algorithm by the Tycho Tracker author, designed for level-0.5 FITS); screened using results pre-run by specialist.py.
[ 0, 0, 0 ]
[[[1.0, 943.0, 741.0], [2.0, 914.0, 724.0], [3.0, 881.0, 709.0], [4.0, 813.0, 674.0], [5.0, 779.0, 654.0], [6.0, 744.0, 641.0], [7.0, 711.0, 625.0], [8.0, 678.0, 607.0]], [[0.0, 356.3, 428.5], [1.0, 350.0, 426.8], [2.0, 343.8, 425.0], [3.0, 337.5, 423.3], [4.0, 325.0, 419.8], [5.0, 318.8, 418.1], [6.0, 312.5, 416.3], [...
0
false
d
Ground-truth self-score 1 (1 both inline and via the track.csv file); independent recomputation (without make_gt, taking raw points by rid from reports.json and matching frames within ±90 s) agrees point by point. gt_check stack_z=3.8; stacked SNR along the ground truth this time = 8.0. Screening: distribution baseline...
All three runs linked points within the CME or the southeast streamer (r=176–488 px) and did not look at the upper-left edge of the field of view.
Enter as benchmark candidate. Still missing: screening by a second model family, and rerunning a genuine single-frame screen and Qwen screen (this source ran idle on the file answer).
End of preview. Expand in Data Studio

Science-Index Video: hard set

78 quantitative questions on raw scientific videos and image sequences from 9 published datasets. Each question gives the solver the raw recording and asks for a number, a set of frame indices, coordinates or a trajectory that can be checked automatically against the original authors' annotation. These are the video questions of Science-Index that remained hard after the full screening and a per-item audit.

Science-Index does not collect new data or new labels. Every answer is derived by a script (make_gt.py, included) from the annotation, catalogue or simulation output that the dataset's authors published.

Sources

Source Dataset Domain Data Items Question types Card
ap-v01 SOHO/LASCO C2/C3 coronal mass ejections Solar physics Real coronagraph image sequences 16 ht_curve, kinematics sources/ap-v01.md
ap-v03 SOHO/LASCO sungrazing comets (Sungrazer Project) Solar system astronomy Real coronagraph image sequences 9 traj_one sources/ap-v03.md
ap-v21 2nd AnDi Challenge, video track Biophysics (single-particle tracking) Simulated fluorescence movies 20 cp_multi, cp_states sources/ap-v21.md
ee-v05 VGC2020 video gauging challenge Hydrology (river surface velocimetry) Real river videos 1 discharge sources/ee-v05.md
ee-v11 Hansbreen glacier calving time-lapse, 2016 Glaciology Real time-lapse photographs 19 big_pairs_12h sources/ee-v11.md
ee-v15 Caltech Fish Counting Fisheries ecology (sonar) Real imaging-sonar videos 4 passage_tracks sources/ee-v15.md
en-v01 Pool boiling high-speed video (UARK-NED3) Thermal engineering Real high-speed videos 1 annot_frame_match sources/en-v01.md
ls-v07 LTDB leukocyte tracking database Cell biology Real microscopy time-lapse 6 traj sources/ls-v07.md
ls-v14 VolPy voltage imaging Neuroscience Real voltage-imaging movies 2 seg_all sources/ls-v14.md

Each card in sources/ describes the original data, where the ground truth comes from, the question types, how items were sampled and known limits.

How an item is built

  1. Pick a published dataset that has raw video (or an image sequence) and the authors' own annotation of a measurable quantity.
  2. Write question templates that ask for that quantity directly from the raw data (counts, event times, positions, trajectories, rates), with the coordinate conventions, units and tolerance stated in the question.
  3. Enumerate instances from the annotation, cut or convert the raw data into input/, and compute the answer with make_gt.py. The answer, unit and tolerance are in task.json; gt/ holds the answer file and, where used, a record of the source rows.

How "hard" was decided

Items passed through, in order: a ground-truth self-check; a distribution baseline (the median or mode of the other items of the same template); a text-only guess from the question alone; a one-step specialist tool or pretrained model for the domain; a look at the first input file only. Items solved by any of these were removed. The rest were run with Claude Opus 5.5 (low effort) as a tool-using agent in an offline sandbox, three times; a question-only (blind) run checked for recall. Every remaining item was then audited by hand: the ground truth was recomputed independently, alternative readings of the question were tested, and the agent's traces were read. Only items audited as genuinely hard (ground truth verified, question unambiguous, agent wrong for a visible reason) are in this set. The mean agent score per item over the three runs is at most 0.713 (median 0.306).

Layout

Path What
data/items.jsonl One row per question with the fields below.
items/<source>/<id>/input/ The raw input the solver gets (video, image frames, TIFF stacks).
items/<source>/<id>/task.json Question, answer, unit, tolerance, ground-truth source and the audit record.
items/<source>/<id>/make_gt.py, gt/ The script that derives the answer from the authors' files, and its output.
sources/<source>.md, sources.csv One card per data source.
scoring/score.py, scoring/structured.py The scorer: scalars within tolerance; structured answers (point sets, trajectories, polylines, time series) by the per-key rules in the code.

Fields of data/items.jsonl

Field Meaning
id, source_id, source_dataset, domain, template_id Identity of the item and its source.
question Full question as given to the solver, including file descriptions, conventions and the answer format.
answer, answer_format, unit, tolerance Ground truth and how it is scored (answer and tolerance are JSON text).
item_dir, input_files, input_bytes, n_materials Where the item's files are.
gt_type, gt_source, provenance, make_gt Kind of ground truth, the exact authors' files and rows it comes from, URLs of the raw data, and the script.
prep_notes How the raw data was cut or converted into input/.
expected_approach, why_multi_look What a solver has to do, and why one look at the data is not enough.
shortcut_risk, specialist_check Shortcuts considered when the item was made, and the specialist-tool baseline.
opus_scores, opus_answers Scores and answers of the three agent runs.
specialist_score, blind_correct Specialist one-step score; whether the question-only run was correct.
audit_category, audit_evidence, why_agent_failed, audit_action Result of the per-item audit (d = genuinely hard).

Notes

  • The ap-v21 inputs are simulated fluorescence movies from the AnDi Challenge benchmark set; all other sources are real recordings.
  • Several ee-v11 questions share the same 12-hour frame window; the audit suggests keeping one question per window when the set is used for evaluation.
  • audit_action lists checks the audit asked for that are still open for some items: for ap-v01, the specialist-tool and one-look screens are to be rerun (they failed silently on these items); for ee-v05, a blind run and a second model family.
  • Each item keeps the license and terms of its source dataset (see sources.csv and the cards).
Downloads last month
73