Dataset Viewer
Auto-converted to Parquet Duplicate
model_group
stringclasses
3 values
model_or_sampler
stringlengths
6
12
semi_ar_dagger_in_table
stringclasses
2 values
HumanEval
float64
15.2
78.1
MBPP
float64
23
62.8
GSM8K
float64
48.8
83.9
MATH500
float64
7.2
64.2
GPQA
float64
26.1
32.8
Countdown
float64
7.7
36.3
Sudoku
float64
0
33.6
reported_average
float64
21.3
45.3
Autoregressive LLMs
LLaMA-3.1-8B
no
53.1
56.7
83.9
23.8
31
27
0
39.4
Autoregressive LLMs
Mistral-7B
no
43.9
37
49.4
7.2
28.1
22.7
0
26.9
Autoregressive LLMs
Qwen-2.5-7B
no
78.1
62.8
71.9
64.2
32.8
7.7
0
45.3
LLaDA-8B-Instruct
Uniform
no
15.2
24.6
48.8
15
29
14.4
2.2
21.3
LLaDA-8B-Instruct
Confidence
no
27.4
42.4
59.1
20.8
27.9
34
23.8
33.6
LLaDA-8B-Instruct
Entropy
no
28.1
42.2
60.9
11.2
28.4
33.8
1.6
29.4
LLaDA-8B-Instruct
Margin
no
32.3
42.4
58.3
19.8
28.4
33.9
26.6
34.5
LLaDA-8B-Instruct
EB-Sampler
no
26.8
43.3
61.2
11.6
29.5
34.1
24.2
33
LLaDA-8B-Instruct
Semi-AR
yes
39
45.2
77.9
27.6
27.7
32.6
0
35.7
LLaDA-8B-Instruct
Fast-dLLM
yes
35.4
44.7
78.2
28.4
28.6
23.6
0
34.1
LLaDA-8B-Instruct
UNCODE
no
42.1
47.8
79.2
34.8
29.2
36.3
29.8
42.7
LLaDA-1.5-8B
Uniform
no
17.7
23
52.7
20
28.1
15.8
3.4
23
LLaDA-1.5-8B
Confidence
no
28.1
43.3
60.7
22.8
28.7
33.8
24.8
34.6
LLaDA-1.5-8B
Entropy
no
32.9
44
60.3
11.2
26.6
34.7
0.2
30
LLaDA-1.5-8B
Margin
no
25
43.3
57.5
23.2
28.4
31.8
33.6
34.7
LLaDA-1.5-8B
EB-Sampler
no
32.9
43.6
61.1
13.4
26.6
34.6
0.2
30.3
LLaDA-1.5-8B
Semi-AR
yes
39.6
46.8
80.7
34.2
26.1
32.4
0
37.1
LLaDA-1.5-8B
Fast-dLLM
yes
37.2
46.1
80.8
31.2
27.9
23.6
0
36.7
LLaDA-1.5-8B
UNCODE
no
46.3
49.9
82.2
37.4
28.8
35
33.4
44.7

Masked diffusion decoder scores on seven benchmarks

Author-prepared CSV transcription of the published results in UNCODE. The release contains 19 model or decoder configurations on seven benchmark tasks.

Reading the columns

Each CSV row identifies a reference model or a sampler evaluated on a particular model. Filter by model_group to compare unmasking strategies on the same backbone. semi_ar_dagger_in_table preserves the footnote marker in the original table. The last column is the average printed by the authors.

The seven task columns are HumanEval, MBPP, GSM8K, MATH500, GPQA, Countdown and Sudoku. The table caption specifies the number of shots separately for each task. Dream-7B results are reported separately in the paper's appendix. This file covers Table 1 in full.

Source and citation

The measurements come from the original paper. Cite the paper when using the results.

Downloads last month
13

Paper for chengpingan/masked-diffusion-decoder-results