File size: 3,441 Bytes
7163353
 
 
 
 
 
 
 
 
 
835a8c4
 
 
 
7163353
 
 
 
 
 
 
 
 
 
 
7e06c1d
7163353
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
835a8c4
 
 
 
 
 
 
 
 
 
 
 
 
7163353
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
---
license: mit
language:
- en
pretty_name: "Same Content, Different Format: worked-example data"
configs:
- config_name: train
  data_files: data/train.jsonl
- config_name: holdout
  data_files: data/holdout.jsonl
- config_name: near_transfer
  data_files: data/near_transfer.jsonl
- config_name: gsm8k_test
  data_files: data/gsm8k_test.jsonl
tags:
- math
- worked-examples
- gsm8k
- continual-pretraining
---

# Same Content, Different Format: worked-example data

Data for the worked-example experiment of *Same Content, Different Format: QA Augmentation and Faded Worked Examples
Under a Fixed Training Budget*. The six trained models are in
[nzhao721/worked-examples-models](https://huggingface.co/nzhao721/worked-examples-models).

## Configs

### `train` (10,000 rows)

The training instances: 5,000 GSM8K training families with two instances each (3,333 answer-augmented, 3,333
rephrased, 1,667 self-verification and 1,667 FOBAR), written by Claude Sonnet 5.5 with MetaMathQA's prompts and
generation code (Yu et al., 2024; commit fe667b14).

| Field | Meaning |
|---|---|
| `instance_id`, `family_id` | instance and family (one family = all rewrites of one GSM8K seed question) |
| `type` | `GSM_AnsAug`, `GSM_Rephrased`, `GSM_SV` or `GSM_FOBAR` |
| `question`, `steps`, `final_answer` | the training question, its step-by-step solution and its final answer |
| `seed_question` | the original GSM8K question of the family |
| `writer_model` | the model that wrote the instance |
| `judge_sonnet`, `judge_argon` | the two LLM judges' audits of the solution (Claude Sonnet 5.5 and Gemini 4 Argon) |

Every instance's final answer matches the gold answer, and neither judge flagged a reasoning error or a slip.

### `holdout` (5,606 rows)

The evaluation holdout: every original question plus every rewrite whose verifying answer matched the gold and that
both judges accepted, 3,247 GSM8K and 2,359 MATH prompts across 1,488 families. No family is shared with training.
`prompt` is the direct-condition prompt (`Problem: q\nAnswer:`); the open condition uses the problem alone.
`final_answer` is the full gold answer.

### `near_transfer` (1,667 rows)

The original GSM8K question of every training family that has no answer-augmented instance. Training showed only
rephrased, self-verification or backward versions of these problems, never the original wording or its solution. A
rephrased training instance has the same final answer as its original question.

### `gsm8k_test` (1,319 rows)

The GSM8K test split (Cobbe et al., 2021), in the same schema, with the final answer taken from each solution's
`####` line. No test problem belongs to a training or holdout family.

All evaluation configs share the holdout's schema.

## Tokenized training streams (`tokenized/`)

The exact token streams each arm was trained on, one folder per arm (`bare`, `complete`, `fade_ordered`,
`fade_shuffled`, `bare_pm`, `complete_pm`): `shard-00000.npy` (token ids, GPT-NeoX OLMo tokenizer, packed into
2,048-token windows in training order), `label_mask-00000.npy` (1 where a token carries loss) and `manifest.json`.
`tokenized/pack_summary.json` gives each arm's windows, tokens, loss tokens and passes. The two fade arms draw on the
same 50,000 documents in different orders.

## Licensing

GSM8K and MATH are released under the MIT license; MetaMathQA's prompts and code are used as released. The rewrites
were written by Claude Sonnet 5.5.