Datasets:
Add the near-transfer and GSM8K test evaluation sets
Browse files- README.md +17 -0
- data/gsm8k_test.jsonl +0 -0
- data/near_transfer.jsonl +0 -0
README.md
CHANGED
|
@@ -8,6 +8,10 @@ configs:
|
|
| 8 |
data_files: data/train.jsonl
|
| 9 |
- config_name: holdout
|
| 10 |
data_files: data/holdout.jsonl
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
tags:
|
| 12 |
- math
|
| 13 |
- worked-examples
|
|
@@ -47,6 +51,19 @@ both judges accepted, 3,247 GSM8K and 2,359 MATH prompts across 1,488 families.
|
|
| 47 |
`prompt` is the direct-condition prompt (`Problem: q\nAnswer:`); the open condition uses the problem alone.
|
| 48 |
`final_answer` is the full gold answer.
|
| 49 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
## Tokenized training streams (`tokenized/`)
|
| 51 |
|
| 52 |
The exact token streams each arm was trained on, one folder per arm (`bare`, `complete`, `fade_ordered`,
|
|
|
|
| 8 |
data_files: data/train.jsonl
|
| 9 |
- config_name: holdout
|
| 10 |
data_files: data/holdout.jsonl
|
| 11 |
+
- config_name: near_transfer
|
| 12 |
+
data_files: data/near_transfer.jsonl
|
| 13 |
+
- config_name: gsm8k_test
|
| 14 |
+
data_files: data/gsm8k_test.jsonl
|
| 15 |
tags:
|
| 16 |
- math
|
| 17 |
- worked-examples
|
|
|
|
| 51 |
`prompt` is the direct-condition prompt (`Problem: q\nAnswer:`); the open condition uses the problem alone.
|
| 52 |
`final_answer` is the full gold answer.
|
| 53 |
|
| 54 |
+
### `near_transfer` (1,667 rows)
|
| 55 |
+
|
| 56 |
+
The original GSM8K question of every training family that has no answer-augmented instance. Training showed only
|
| 57 |
+
rephrased, self-verification or backward versions of these problems, never the original wording or its solution. A
|
| 58 |
+
rephrased training instance has the same final answer as its original question.
|
| 59 |
+
|
| 60 |
+
### `gsm8k_test` (1,319 rows)
|
| 61 |
+
|
| 62 |
+
The GSM8K test split (Cobbe et al., 2021), in the same schema, with the final answer taken from each solution's
|
| 63 |
+
`####` line. No test problem belongs to a training or holdout family.
|
| 64 |
+
|
| 65 |
+
All evaluation configs share the holdout's schema.
|
| 66 |
+
|
| 67 |
## Tokenized training streams (`tokenized/`)
|
| 68 |
|
| 69 |
The exact token streams each arm was trained on, one folder per arm (`bare`, `complete`, `fade_ordered`,
|
data/gsm8k_test.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/near_transfer.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|