hanya256 commited on
Commit
b835b05
·
verified ·
1 Parent(s): 71654fd

Release four archived main MATH experiments

Browse files
README.md ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: DIBA archived main MATH experiments
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ ---
8
+
9
+ # Archived main MATH experiments
10
+
11
+ This release preserves the four original MATH runs (GRPO/DAPO × 3B/7B) from the experiment archive. It includes the saved metrics here and the generated target responses and sample features in [RLMIA/DIBA-MATH-RESULTS](https://huggingface.co/datasets/RLMIA/DIBA-MATH-RESULTS).
12
+
13
+ Each run contains 1,269 member and 1,269 nonmember prompts, with 8 target responses per prompt. The archive directories record temperature **0.5**. These are archived runs, separate from the default temperature 0.7 of the portable runner.
14
+
15
+ ## Download generated responses
16
+
17
+ ```bash
18
+ hf download RLMIA/DIBA-MATH-RESULTS --repo-type dataset --local-dir runs/released-main
19
+ gzip -dk runs/released-main/grpo-3b/responses_ft_member.jsonl.gz
20
+ ```
21
+
22
+ The dataset contains only the four main runs. Model weights are in their model repositories. Tuning runs, ablations, intermediate checkpoints, and training logs are excluded.
23
+
24
+ ## Contents
25
+
26
+ Each run has the original `metrics_prompt.npz`, `evaluation_summary.csv`, and the unified/specialist reports. The Hugging Face dataset additionally contains compressed target-response JSONL files for members/nonmembers and `computed_features.jsonl.gz`. [manifest.json](manifest.json) records hashes, row counts, feature shape, and non-finite counts.
27
+
28
+ The original archive contains **no reference-response JSONL files**. It does contain the saved reference scores in both the sample feature cache and prompt feature cache. Recomputing verifier scores from both models requires generating the missing reference responses with the corresponding original settings.
29
+
30
+ ## Feature layout and evaluation
31
+
32
+ The archived `feats` array has five columns, in this order:
33
+
34
+ 1. `neg_log_likelihood`
35
+ 2. `zlib_entropy_ratio`
36
+ 3. `reverse_kl_div`
37
+ 4. `avg_base_score`
38
+ 5. `avg_ft_score`
39
+
40
+ The last two columns are verifier scores. This is a different layout from the portable runner's seven-column output. Use the evaluator that consumes the saved numeric array directly:
41
+
42
+ ```bash
43
+ python scripts/evaluate_cached.py \
44
+ --input runs/released-main/grpo-3b/metrics_prompt.npz \
45
+ --out_dir runs/released-grpo-3b-reeval
46
+ ```
47
+
48
+ This fits the current unified predictor on the saved feature matrix; original reports remain unchanged. Add `--limit_per_class 8` for an execution check only.
49
+
50
+ Original values are preserved: DAPO-3B contains one non-finite detector-feature entry and GRPO-7B contains two. The raw sample-feature caches also contain non-finite values in unused features. The evaluator reports how many detector-feature entries it replaces with zero, following the current evaluation pipeline; it does not modify the archive. The historical reports and a new evaluation should be identified separately.
dapo-3b/computed_features.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81cd45cfa442d171501d60bb4f23b21e11faf3d64b0e156e9530546a174a4a15
3
+ size 1155442
dapo-3b/evaluation_summary.csv ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ Split,GRPO,,
2
+ ,# Samples,AUC,TPR @ 1.0% FPR,TPR @ 0.1% FPR
3
+ ALL-0,95,0.559,0.000,0.000
4
+ ALL-1,118,0.518,0.014,0.014
5
+ Residual,294,0.912,0.333,0.282
6
+ Unified,508,0.834,0.213,0.154
dapo-3b/metrics_prompt.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f29e9869eca6c5b7acad46a0a25b38df2e7ddf6d306d6ca87b861e5818ba1384
3
+ size 265236
dapo-3b/report_specialist_models_combined.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Combined Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.7703 \u00b1 0.0194",
4
+ "f1_score": "0.7814 \u00b1 0.0182",
5
+ "auc": "0.8399 \u00b1 0.0188",
6
+ "tpr_at_0.1%_fpr": "0.1639 \u00b1 0.0438",
7
+ "tpr_at_1.0%_fpr": "0.2222 \u00b1 0.0381",
8
+ "tpr_at_5.0%_fpr": "0.3845 \u00b1 0.0607"
9
+ }
dapo-3b/report_specialist_models_individual.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --- Individual Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "all-0": {
4
+ "accuracy": "0.7002 \u00b1 0.0525",
5
+ "f1_score": "0.0000 \u00b1 0.0000",
6
+ "auc": "0.5595 \u00b1 0.0526",
7
+ "tpr_at_0.1%_fpr": "0.0000 \u00b1 0.0000",
8
+ "tpr_at_1.0%_fpr": "0.0000 \u00b1 0.0000",
9
+ "tpr_at_5.0%_fpr": "0.1070 \u00b1 0.0561"
10
+ },
11
+ "all-1": {
12
+ "accuracy": "0.6430 \u00b1 0.0440",
13
+ "f1_score": "0.7819 \u00b1 0.0329",
14
+ "auc": "0.5180 \u00b1 0.0306",
15
+ "tpr_at_0.1%_fpr": "0.0143 \u00b1 0.0136",
16
+ "tpr_at_1.0%_fpr": "0.0143 \u00b1 0.0136",
17
+ "tpr_at_5.0%_fpr": "0.0376 \u00b1 0.0234"
18
+ },
19
+ "residual": {
20
+ "accuracy": "0.8425 \u00b1 0.0186",
21
+ "f1_score": "0.8510 \u00b1 0.0172",
22
+ "auc": "0.9117 \u00b1 0.0206",
23
+ "tpr_at_0.1%_fpr": "0.2818 \u00b1 0.0854",
24
+ "tpr_at_1.0%_fpr": "0.3331 \u00b1 0.0847",
25
+ "tpr_at_5.0%_fpr": "0.5425 \u00b1 0.0727"
26
+ }
27
+ }
dapo-3b/report_unified_model.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Unified Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.7620 \u00b1 0.0227",
4
+ "f1_score": "0.7692 \u00b1 0.0203",
5
+ "auc": "0.8342 \u00b1 0.0236",
6
+ "tpr_at_0.1%_fpr": "0.1544 \u00b1 0.0437",
7
+ "tpr_at_1.0%_fpr": "0.2135 \u00b1 0.0413",
8
+ "tpr_at_5.0%_fpr": "0.3727 \u00b1 0.0503"
9
+ }
dapo-3b/responses_ft_member.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e949750e073d27f910000808e5751f6c45de4a393734993cc364e3fa683d55d
3
+ size 2112436
dapo-3b/responses_ft_nonmember.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d79ebd1d25be9c2fb97c3a52d094163ea1be8f98b4914c5e04a56d65c3ab2a0a
3
+ size 2609915
dapo-7b/computed_features.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:877464d05f1fb9b0fa77e6581b43785f69d45c92575375a183fbba5f64ee0ee5
3
+ size 1154691
dapo-7b/evaluation_summary.csv ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ Split,GRPO,,
2
+ ,# Samples,AUC,TPR @ 1.0% FPR,TPR @ 0.1% FPR
3
+ ALL-0,69,0.528,0.030,0.030
4
+ ALL-1,189,0.485,0.016,0.016
5
+ Residual,250,0.937,0.354,0.280
6
+ Unified,508,0.819,0.214,0.107
dapo-7b/metrics_prompt.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:193ddc92576c16afa675bb085bb5f5cc82f6cd6be3310e25776ea61e4df57855
3
+ size 265236
dapo-7b/report_specialist_models_combined.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Combined Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.7431 \u00b1 0.0159",
4
+ "f1_score": "0.7717 \u00b1 0.0150",
5
+ "auc": "0.8228 \u00b1 0.0163",
6
+ "tpr_at_0.1%_fpr": "0.1372 \u00b1 0.0797",
7
+ "tpr_at_1.0%_fpr": "0.2317 \u00b1 0.0743",
8
+ "tpr_at_5.0%_fpr": "0.3790 \u00b1 0.0440"
9
+ }
dapo-7b/report_specialist_models_individual.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --- Individual Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "all-0": {
4
+ "accuracy": "0.7218 \u00b1 0.0438",
5
+ "f1_score": "0.0000 \u00b1 0.0000",
6
+ "auc": "0.5282 \u00b1 0.0611",
7
+ "tpr_at_0.1%_fpr": "0.0299 \u00b1 0.0381",
8
+ "tpr_at_1.0%_fpr": "0.0299 \u00b1 0.0381",
9
+ "tpr_at_5.0%_fpr": "0.1008 \u00b1 0.0862"
10
+ },
11
+ "all-1": {
12
+ "accuracy": "0.5974 \u00b1 0.0157",
13
+ "f1_score": "0.7475 \u00b1 0.0128",
14
+ "auc": "0.4854 \u00b1 0.0324",
15
+ "tpr_at_0.1%_fpr": "0.0160 \u00b1 0.0200",
16
+ "tpr_at_1.0%_fpr": "0.0160 \u00b1 0.0200",
17
+ "tpr_at_5.0%_fpr": "0.0477 \u00b1 0.0218"
18
+ },
19
+ "residual": {
20
+ "accuracy": "0.8588 \u00b1 0.0190",
21
+ "f1_score": "0.8587 \u00b1 0.0217",
22
+ "auc": "0.9374 \u00b1 0.0156",
23
+ "tpr_at_0.1%_fpr": "0.2796 \u00b1 0.1529",
24
+ "tpr_at_1.0%_fpr": "0.3542 \u00b1 0.1348",
25
+ "tpr_at_5.0%_fpr": "0.6644 \u00b1 0.0805"
26
+ }
27
+ }
dapo-7b/report_unified_model.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Unified Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.7396 \u00b1 0.0192",
4
+ "f1_score": "0.7539 \u00b1 0.0205",
5
+ "auc": "0.8189 \u00b1 0.0247",
6
+ "tpr_at_0.1%_fpr": "0.1072 \u00b1 0.0728",
7
+ "tpr_at_1.0%_fpr": "0.2136 \u00b1 0.0621",
8
+ "tpr_at_5.0%_fpr": "0.3719 \u00b1 0.0567"
9
+ }
dapo-7b/responses_ft_member.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b040b2c299681715a705f24377dccddaa8e1af2eb31ce64264ada3e0a558637e
3
+ size 1849707
dapo-7b/responses_ft_nonmember.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:232ee6d8537b207a44aa88827e7595256d69104daa3559b58561b554220d1d29
3
+ size 2372046
grpo-3b/computed_features.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d921b7f157c2ef10c2c710258ba37cb058c694d9541c46e2b9bd90fae2d1928a
3
+ size 1169561
grpo-3b/evaluation_summary.csv ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ Split,GRPO,,
2
+ ,# Samples,AUC,TPR @ 1.0% FPR,TPR @ 0.1% FPR
3
+ ALL-0,108,0.687,0.047,0.047
4
+ ALL-1,119,0.551,0.028,0.028
5
+ Residual,281,0.765,0.197,0.130
6
+ Unified,508,0.713,0.123,0.073
grpo-3b/metrics_prompt.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:492a02ba87767dbb4cbd12f62b6fbd196a14891ccd685e7d6f881a12da2ab881
3
+ size 265236
grpo-3b/report_specialist_models_combined.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Combined Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.6718 \u00b1 0.0300",
4
+ "f1_score": "0.6746 \u00b1 0.0360",
5
+ "auc": "0.7275 \u00b1 0.0317",
6
+ "tpr_at_0.1%_fpr": "0.0677 \u00b1 0.0404",
7
+ "tpr_at_1.0%_fpr": "0.1189 \u00b1 0.0351",
8
+ "tpr_at_5.0%_fpr": "0.1954 \u00b1 0.0253"
9
+ }
grpo-3b/report_specialist_models_individual.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --- Individual Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "all-0": {
4
+ "accuracy": "0.6335 \u00b1 0.0357",
5
+ "f1_score": "0.3365 \u00b1 0.0446",
6
+ "auc": "0.6871 \u00b1 0.0209",
7
+ "tpr_at_0.1%_fpr": "0.0470 \u00b1 0.0399",
8
+ "tpr_at_1.0%_fpr": "0.0470 \u00b1 0.0399",
9
+ "tpr_at_5.0%_fpr": "0.1823 \u00b1 0.0391"
10
+ },
11
+ "all-1": {
12
+ "accuracy": "0.6322 \u00b1 0.0407",
13
+ "f1_score": "0.7718 \u00b1 0.0299",
14
+ "auc": "0.5514 \u00b1 0.0228",
15
+ "tpr_at_0.1%_fpr": "0.0275 \u00b1 0.0253",
16
+ "tpr_at_1.0%_fpr": "0.0275 \u00b1 0.0253",
17
+ "tpr_at_5.0%_fpr": "0.0422 \u00b1 0.0308"
18
+ },
19
+ "residual": {
20
+ "accuracy": "0.7025 \u00b1 0.0286",
21
+ "f1_score": "0.6804 \u00b1 0.0341",
22
+ "auc": "0.7649 \u00b1 0.0283",
23
+ "tpr_at_0.1%_fpr": "0.1297 \u00b1 0.0806",
24
+ "tpr_at_1.0%_fpr": "0.1968 \u00b1 0.0817",
25
+ "tpr_at_5.0%_fpr": "0.3134 \u00b1 0.0639"
26
+ }
27
+ }
grpo-3b/report_unified_model.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Unified Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.6596 \u00b1 0.0200",
4
+ "f1_score": "0.6574 \u00b1 0.0232",
5
+ "auc": "0.7128 \u00b1 0.0157",
6
+ "tpr_at_0.1%_fpr": "0.0733 \u00b1 0.0359",
7
+ "tpr_at_1.0%_fpr": "0.1229 \u00b1 0.0139",
8
+ "tpr_at_5.0%_fpr": "0.2135 \u00b1 0.0236"
9
+ }
grpo-3b/responses_ft_member.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1ac260a5ead695a92efa44c2a27a43bb8519874974725d06fbd38aac110ed60
3
+ size 2300633
grpo-3b/responses_ft_nonmember.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e27801fddb92936b837ff6977af9c605e4a4bcf3572be7fddd53648209771ed
3
+ size 2698480
grpo-7b/computed_features.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e7d682fb7c84689add822f370b60de391b5ddebbb47087fe251bde0d5e38a14f
3
+ size 1123901
grpo-7b/evaluation_summary.csv ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ Split,GRPO,,
2
+ ,# Samples,AUC,TPR @ 1.0% FPR,TPR @ 0.1% FPR
3
+ ALL-0,76,0.661,0.064,0.064
4
+ ALL-1,191,0.525,0.005,0.005
5
+ Residual,240,0.877,0.279,0.162
6
+ Unified,508,0.750,0.151,0.084
grpo-7b/metrics_prompt.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e3021fbd1aea63fd5687d3f8ab799d8bd69e13036d5bc51eb188ed9688e62c75
3
+ size 265236
grpo-7b/report_specialist_models_combined.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Combined Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.6970 \u00b1 0.0196",
4
+ "f1_score": "0.7235 \u00b1 0.0204",
5
+ "auc": "0.7694 \u00b1 0.0164",
6
+ "tpr_at_0.1%_fpr": "0.0733 \u00b1 0.0401",
7
+ "tpr_at_1.0%_fpr": "0.1568 \u00b1 0.0241",
8
+ "tpr_at_5.0%_fpr": "0.2506 \u00b1 0.0141"
9
+ }
grpo-7b/report_specialist_models_individual.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --- Individual Specialist Model Performance (5-Fold CV) ---
2
+ {
3
+ "all-0": {
4
+ "accuracy": "0.6497 \u00b1 0.0457",
5
+ "f1_score": "0.1537 \u00b1 0.0482",
6
+ "auc": "0.6610 \u00b1 0.0579",
7
+ "tpr_at_0.1%_fpr": "0.0643 \u00b1 0.0516",
8
+ "tpr_at_1.0%_fpr": "0.0643 \u00b1 0.0516",
9
+ "tpr_at_5.0%_fpr": "0.1706 \u00b1 0.0678"
10
+ },
11
+ "all-1": {
12
+ "accuracy": "0.5879 \u00b1 0.0385",
13
+ "f1_score": "0.7353 \u00b1 0.0304",
14
+ "auc": "0.5254 \u00b1 0.0240",
15
+ "tpr_at_0.1%_fpr": "0.0052 \u00b1 0.0103",
16
+ "tpr_at_1.0%_fpr": "0.0052 \u00b1 0.0103",
17
+ "tpr_at_5.0%_fpr": "0.0409 \u00b1 0.0337"
18
+ },
19
+ "residual": {
20
+ "accuracy": "0.7974 \u00b1 0.0163",
21
+ "f1_score": "0.7855 \u00b1 0.0160",
22
+ "auc": "0.8775 \u00b1 0.0150",
23
+ "tpr_at_0.1%_fpr": "0.1618 \u00b1 0.0875",
24
+ "tpr_at_1.0%_fpr": "0.2793 \u00b1 0.0798",
25
+ "tpr_at_5.0%_fpr": "0.4748 \u00b1 0.0469"
26
+ }
27
+ }
grpo-7b/report_unified_model.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ --- Unified Model Performance (5-Fold CV) ---
2
+ {
3
+ "accuracy": "0.6726 \u00b1 0.0120",
4
+ "f1_score": "0.6865 \u00b1 0.0122",
5
+ "auc": "0.7495 \u00b1 0.0128",
6
+ "tpr_at_0.1%_fpr": "0.0843 \u00b1 0.0468",
7
+ "tpr_at_1.0%_fpr": "0.1513 \u00b1 0.0380",
8
+ "tpr_at_5.0%_fpr": "0.2474 \u00b1 0.0198"
9
+ }
grpo-7b/responses_ft_member.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:197cadef23121b2725eeb911d7c321f33e7f8df430aabe7faf4c5946d9d3cce2
3
+ size 1816268
grpo-7b/responses_ft_nonmember.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:082660e249811ac3c806537a8b818aad968dc4f70d0da89434b95c9e607caadd
3
+ size 2174591
manifest.json ADDED
@@ -0,0 +1,253 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "source_archive": "math.zip",
4
+ "source_archive_sha256": "4864c8b527f89f295631543fe7a967226feb939bf3a31ca3252af4d4afad6846",
5
+ "temperature_from_archive_path": 0.5,
6
+ "num_samples": 8,
7
+ "feature_names": [
8
+ "neg_log_likelihood",
9
+ "zlib_entropy_ratio",
10
+ "reverse_kl_div",
11
+ "avg_base_score",
12
+ "avg_ft_score"
13
+ ],
14
+ "reference_response_files_in_archive": false,
15
+ "runs": [
16
+ {
17
+ "run": "grpo-3b",
18
+ "source_prefix": "math_grpo_3b_temp/t=0.5/",
19
+ "files": [
20
+ {
21
+ "path": "grpo-3b/metrics_prompt.npz",
22
+ "bytes": 265236,
23
+ "sha256": "492a02ba87767dbb4cbd12f62b6fbd196a14891ccd685e7d6f881a12da2ab881"
24
+ },
25
+ {
26
+ "path": "grpo-3b/evaluation_summary.csv",
27
+ "bytes": 181,
28
+ "sha256": "802dbd164465f5bb99af60c0a9fc1f61e181979f9aeb64e5cccbf24281091597"
29
+ },
30
+ {
31
+ "path": "grpo-3b/report_unified_model.txt",
32
+ "bytes": 292,
33
+ "sha256": "71deb7c1a030b2cf8fe1e0a20c25ce7fcf7a041a4ae7cd1865a6b512c17b6b9a"
34
+ },
35
+ {
36
+ "path": "grpo-3b/report_specialist_models_individual.txt",
37
+ "bytes": 884,
38
+ "sha256": "2e349aeb33ff82266235d6ec996d56e5b8037fcb697e117fbfe56fd69eb4945f"
39
+ },
40
+ {
41
+ "path": "grpo-3b/report_specialist_models_combined.txt",
42
+ "bytes": 304,
43
+ "sha256": "b86383ed806df74c420aa0c9547a6dfd055b29f758c36c10147952f5506e8721"
44
+ },
45
+ {
46
+ "path": "grpo-3b/responses_ft_member.jsonl.gz",
47
+ "bytes": 2300633,
48
+ "sha256": "b1ac260a5ead695a92efa44c2a27a43bb8519874974725d06fbd38aac110ed60",
49
+ "uncompressed_sha256": "c9c06967ea04a570521f2e9b5a7c3a7923a536965418660f544c39921e919bd6",
50
+ "records": 10152
51
+ },
52
+ {
53
+ "path": "grpo-3b/responses_ft_nonmember.jsonl.gz",
54
+ "bytes": 2698480,
55
+ "sha256": "4e27801fddb92936b837ff6977af9c605e4a4bcf3572be7fddd53648209771ed",
56
+ "uncompressed_sha256": "aacb52f74d109d83d97b816589a4a7ab57c1c2354139341cb8f9d30080ef5a47",
57
+ "records": 10152
58
+ },
59
+ {
60
+ "path": "grpo-3b/computed_features.jsonl.gz",
61
+ "bytes": 1169561,
62
+ "sha256": "d921b7f157c2ef10c2c710258ba37cb058c694d9541c46e2b9bd90fae2d1928a",
63
+ "uncompressed_sha256": "84b283dae1a8702b9e526104273f477f8da1521dbea4de45207c95c3af55d5e9",
64
+ "records": 20304
65
+ }
66
+ ],
67
+ "feature_shape": [
68
+ 2538,
69
+ 5
70
+ ],
71
+ "nonfinite_feature_entries": 0,
72
+ "member_prompts": 1269,
73
+ "nonmember_prompts": 1269
74
+ },
75
+ {
76
+ "run": "grpo-7b",
77
+ "source_prefix": "math_grpo_7b_temp/t=0.5/",
78
+ "files": [
79
+ {
80
+ "path": "grpo-7b/metrics_prompt.npz",
81
+ "bytes": 265236,
82
+ "sha256": "e3021fbd1aea63fd5687d3f8ab799d8bd69e13036d5bc51eb188ed9688e62c75"
83
+ },
84
+ {
85
+ "path": "grpo-7b/evaluation_summary.csv",
86
+ "bytes": 180,
87
+ "sha256": "dc47f778e2b615a9bcddf698140f598b8f4c863d456566575249f9b5e2ae8a4d"
88
+ },
89
+ {
90
+ "path": "grpo-7b/report_unified_model.txt",
91
+ "bytes": 292,
92
+ "sha256": "8146d5640756904b8249c4f23d0f74f8e525724b18e11ac8d36c66ba7b3265e9"
93
+ },
94
+ {
95
+ "path": "grpo-7b/report_specialist_models_individual.txt",
96
+ "bytes": 884,
97
+ "sha256": "b29a557c293712423482e10b9f9759f0e6cca73e6f12c21511b25a36bf716d59"
98
+ },
99
+ {
100
+ "path": "grpo-7b/report_specialist_models_combined.txt",
101
+ "bytes": 304,
102
+ "sha256": "38b9523a479129d64d232dd44ce1b837309c8bac61dc4ac435f4a78b1f059c86"
103
+ },
104
+ {
105
+ "path": "grpo-7b/responses_ft_member.jsonl.gz",
106
+ "bytes": 1816268,
107
+ "sha256": "197cadef23121b2725eeb911d7c321f33e7f8df430aabe7faf4c5946d9d3cce2",
108
+ "uncompressed_sha256": "3b077d53f9df8bd05cc454f068ea66abfadcafc184461c5ef6710976c3ef05aa",
109
+ "records": 10152
110
+ },
111
+ {
112
+ "path": "grpo-7b/responses_ft_nonmember.jsonl.gz",
113
+ "bytes": 2174591,
114
+ "sha256": "082660e249811ac3c806537a8b818aad968dc4f70d0da89434b95c9e607caadd",
115
+ "uncompressed_sha256": "f8567ed9cf1ac93e6cb64a912188d0d45f16365e83faf3793764efcbd5d0baf4",
116
+ "records": 10152
117
+ },
118
+ {
119
+ "path": "grpo-7b/computed_features.jsonl.gz",
120
+ "bytes": 1123901,
121
+ "sha256": "e7d682fb7c84689add822f370b60de391b5ddebbb47087fe251bde0d5e38a14f",
122
+ "uncompressed_sha256": "716bd75d3e573cc45ea98bd9303b38a6bcb10d3b3503c35eaa003c5e05a271ca",
123
+ "records": 20304
124
+ }
125
+ ],
126
+ "feature_shape": [
127
+ 2538,
128
+ 5
129
+ ],
130
+ "nonfinite_feature_entries": 2,
131
+ "member_prompts": 1269,
132
+ "nonmember_prompts": 1269
133
+ },
134
+ {
135
+ "run": "dapo-3b",
136
+ "source_prefix": "math_dapo_3b_temp/t=0.5/",
137
+ "files": [
138
+ {
139
+ "path": "dapo-3b/metrics_prompt.npz",
140
+ "bytes": 265236,
141
+ "sha256": "f29e9869eca6c5b7acad46a0a25b38df2e7ddf6d306d6ca87b861e5818ba1384"
142
+ },
143
+ {
144
+ "path": "dapo-3b/evaluation_summary.csv",
145
+ "bytes": 180,
146
+ "sha256": "16f70246041b0b9ba9cdfd245ab582a4b003363a49d1e80670227dad0c925e52"
147
+ },
148
+ {
149
+ "path": "dapo-3b/report_unified_model.txt",
150
+ "bytes": 292,
151
+ "sha256": "0834f6823f15b26bdaab38dbce5b5d0e091d4ab33e923f59b035e4a6aa36230c"
152
+ },
153
+ {
154
+ "path": "dapo-3b/report_specialist_models_individual.txt",
155
+ "bytes": 884,
156
+ "sha256": "25eb08abb9995da49955f1a02038ad46f732c81d1e07efcba7d8293201c83df2"
157
+ },
158
+ {
159
+ "path": "dapo-3b/report_specialist_models_combined.txt",
160
+ "bytes": 304,
161
+ "sha256": "e086ea55255b69b32131c84b7c45c7d0a767ba6bd79917b1d2948ab33ef507f5"
162
+ },
163
+ {
164
+ "path": "dapo-3b/responses_ft_member.jsonl.gz",
165
+ "bytes": 2112436,
166
+ "sha256": "6e949750e073d27f910000808e5751f6c45de4a393734993cc364e3fa683d55d",
167
+ "uncompressed_sha256": "6913703a8cfd5ee8d5f542578f227d795f4b1b813cac52809024eab8b10c303c",
168
+ "records": 10152
169
+ },
170
+ {
171
+ "path": "dapo-3b/responses_ft_nonmember.jsonl.gz",
172
+ "bytes": 2609915,
173
+ "sha256": "d79ebd1d25be9c2fb97c3a52d094163ea1be8f98b4914c5e04a56d65c3ab2a0a",
174
+ "uncompressed_sha256": "2d7d47220a9f7a3897471ddd90bfeb68c8269d92875ab842d4f042cb5246ea15",
175
+ "records": 10152
176
+ },
177
+ {
178
+ "path": "dapo-3b/computed_features.jsonl.gz",
179
+ "bytes": 1155442,
180
+ "sha256": "81cd45cfa442d171501d60bb4f23b21e11faf3d64b0e156e9530546a174a4a15",
181
+ "uncompressed_sha256": "1af4e2afda720507947a1deaaac38b708b167ffae790d4b3e9c4ad7085d638b2",
182
+ "records": 20304
183
+ }
184
+ ],
185
+ "feature_shape": [
186
+ 2538,
187
+ 5
188
+ ],
189
+ "nonfinite_feature_entries": 1,
190
+ "member_prompts": 1269,
191
+ "nonmember_prompts": 1269
192
+ },
193
+ {
194
+ "run": "dapo-7b",
195
+ "source_prefix": "math_dapo_7b_temp/t=0.5/",
196
+ "files": [
197
+ {
198
+ "path": "dapo-7b/metrics_prompt.npz",
199
+ "bytes": 265236,
200
+ "sha256": "193ddc92576c16afa675bb085bb5f5cc82f6cd6be3310e25776ea61e4df57855"
201
+ },
202
+ {
203
+ "path": "dapo-7b/evaluation_summary.csv",
204
+ "bytes": 180,
205
+ "sha256": "d4cd74663b2293b183aa6ab5407dcaf290f6b0fe354d0b34fcbab21cf82e6987"
206
+ },
207
+ {
208
+ "path": "dapo-7b/report_unified_model.txt",
209
+ "bytes": 292,
210
+ "sha256": "69cbabc5729cd15b5f5778c7144e1e6ead783dce1fd89f72864e61232c0a3456"
211
+ },
212
+ {
213
+ "path": "dapo-7b/report_specialist_models_individual.txt",
214
+ "bytes": 884,
215
+ "sha256": "41df44f3305f8b98102a8325a03bbf072f098b89414058353406f85b52c5fc20"
216
+ },
217
+ {
218
+ "path": "dapo-7b/report_specialist_models_combined.txt",
219
+ "bytes": 304,
220
+ "sha256": "f7990c2e745cedcaf8eba0d0555c2e6f98ad4156df2d4544a691b0be42662338"
221
+ },
222
+ {
223
+ "path": "dapo-7b/responses_ft_member.jsonl.gz",
224
+ "bytes": 1849707,
225
+ "sha256": "b040b2c299681715a705f24377dccddaa8e1af2eb31ce64264ada3e0a558637e",
226
+ "uncompressed_sha256": "88c6ea65d1cb02c3e305b5a967d541ba3886fe202c99ace57b4e9ee47d8b30c8",
227
+ "records": 10152
228
+ },
229
+ {
230
+ "path": "dapo-7b/responses_ft_nonmember.jsonl.gz",
231
+ "bytes": 2372046,
232
+ "sha256": "232ee6d8537b207a44aa88827e7595256d69104daa3559b58561b554220d1d29",
233
+ "uncompressed_sha256": "dcf0478a512f0b513f81eb57c01b2cb65f926270b39149a1eed7bf220db6a879",
234
+ "records": 10152
235
+ },
236
+ {
237
+ "path": "dapo-7b/computed_features.jsonl.gz",
238
+ "bytes": 1154691,
239
+ "sha256": "877464d05f1fb9b0fa77e6581b43785f69d45c92575375a183fbba5f64ee0ee5",
240
+ "uncompressed_sha256": "64d772c16657249dd663d8470292da8162a5316025868b0fcb8bd97af8a808af",
241
+ "records": 20304
242
+ }
243
+ ],
244
+ "feature_shape": [
245
+ 2538,
246
+ 5
247
+ ],
248
+ "nonfinite_feature_entries": 0,
249
+ "member_prompts": 1269,
250
+ "nonmember_prompts": 1269
251
+ }
252
+ ]
253
+ }