nzhao721 commited on
Commit
835a8c4
·
verified ·
1 Parent(s): 7e06c1d

Add the near-transfer and GSM8K test evaluation sets

Browse files
Files changed (3) hide show
  1. README.md +17 -0
  2. data/gsm8k_test.jsonl +0 -0
  3. 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