ClaudioSavelli commited on
Commit
edb1663
·
verified ·
1 Parent(s): 81a9954

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +57 -43
README.md CHANGED
@@ -1,45 +1,59 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: id
5
- dtype: string
6
- - name: question
7
- dtype: string
8
- - name: choices
9
- sequence: string
10
- - name: answer
11
- dtype: int64
12
- - name: topic
13
- dtype: string
14
- - name: subject
15
- dtype: string
16
- - name: hint
17
- dtype: string
18
- - name: input
19
- dtype: string
20
- - name: target
21
- dtype: string
22
- - name: source_index
23
- dtype: int64
24
- splits:
25
- - name: train
26
- num_bytes: 4095164
27
- num_examples: 6508
28
- - name: validation
29
- num_bytes: 1336576
30
- num_examples: 2144
31
- - name: test
32
- num_bytes: 1419479
33
- num_examples: 2224
34
- download_size: 3140720
35
- dataset_size: 6851219
36
- configs:
37
- - config_name: default
38
- data_files:
39
- - split: train
40
- path: data/train-*
41
- - split: validation
42
- path: data/validation-*
43
- - split: test
44
- path: data/test-*
45
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ license: cc-by-sa-4.0
5
+ task_categories:
6
+ - question-answering
7
+ pretty_name: Text-only ScienceQA for Sequential Unlearning
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  ---
9
+
10
+ # Text-only ScienceQA for sequential unlearning
11
+
12
+ This private dataset contains every image-free example from
13
+ `derek-thomas/ScienceQA` at immutable revision
14
+ `f18b0a70359ebfb41f658fd564208d0355b013f4`. The filter is exactly `image is null`, matching
15
+ the ScienceQA preprocessing released with *On Large Language Model Continual
16
+ Unlearning*. Original train, validation, and test membership is preserved.
17
+
18
+ No deletion labels or fixed deletion plan are stored. The paper's reference
19
+ sequence is `biology -> physics -> chemistry -> economics`; consumers should
20
+ filter the lowercase `topic` column at runtime.
21
+
22
+ ## Schema
23
+
24
+ - `id`: deterministic content-derived ID. The Hugging Face source omits the
25
+ original ScienceQA PID.
26
+ - `question`, `choices`, `answer`, `topic`, `subject`, `hint`: normalized source fields.
27
+ - `input`: context (when present), question, and lettered options; never the answer.
28
+ - `target`: `The answer is <LETTER>`.
29
+ - `source_index`: row index in the unfiltered upstream split.
30
+
31
+ Explanations (`lecture` and `solution`) and images are deliberately excluded.
32
+
33
+ ## Counts
34
+
35
+ | Split | Rows |
36
+ |---|---:|
37
+ | train | 6508 |
38
+ | validation | 2144 |
39
+ | test | 2224 |
40
+
41
+ | Reference topic | Train | Validation | Test |
42
+ |---|---:|---:|---:|
43
+ | biology | 1192 | 406 | 397 |
44
+ | physics | 595 | 226 | 192 |
45
+ | chemistry | 403 | 136 | 119 |
46
+ | economics | 237 | 63 | 84 |
47
+
48
+ ## Loading
49
+
50
+ ```python
51
+ from datasets import load_dataset
52
+
53
+ dataset = load_dataset("ClaudioSavelli/ScienceQA", token=True)
54
+ biology_train = dataset["train"].filter(lambda row: row["topic"] == "biology")
55
+ ```
56
+
57
+ The preparation seed is recorded for reproducibility, although dataset creation
58
+ does not use random sampling. See `scienceqa_metadata.json` for content digests
59
+ and the complete per-topic distribution.