Datasets:
Publish complete base, GRPO step140 and RLTT step140 pass@8 artifacts
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +21 -0
- README.md +76 -0
- SHA256SUMS +171 -0
- comparison.json +342 -0
- evaluations/base_pass8_16k_seed42/code/rltt_baseline/eval_math_suite.py +390 -0
- evaluations/base_pass8_16k_seed42/code/scripts/resume_math_eval.py +112 -0
- evaluations/base_pass8_16k_seed42/code/scripts/run_math_eval_suite.py +131 -0
- evaluations/base_pass8_16k_seed42/code/tests/test_math_eval_suite.py +106 -0
- evaluations/base_pass8_16k_seed42/data/questions.jsonl +0 -0
- evaluations/base_pass8_16k_seed42/data/sources.json +91 -0
- evaluations/base_pass8_16k_seed42/eval_exit_code +1 -0
- evaluations/base_pass8_16k_seed42/logs/coordinator.log +9 -0
- evaluations/base_pass8_16k_seed42/logs/rank0.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank1.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank2.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank3.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank4.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank5.log +0 -0
- evaluations/base_pass8_16k_seed42/logs/rank6.log +0 -0
- evaluations/base_pass8_16k_seed42/preflight/code_hashes.json +6 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank0.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank1.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank2.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank3.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank4.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank5.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/engine_rank6.json +32 -0
- evaluations/base_pass8_16k_seed42/preflight/launch_receipt.json +78 -0
- evaluations/base_pass8_16k_seed42/preflight/resume.json +82 -0
- evaluations/base_pass8_16k_seed42/protocol.json +47 -0
- evaluations/base_pass8_16k_seed42/provenance/parent_protocol.json +44 -0
- evaluations/base_pass8_16k_seed42/responses/rank0.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank1.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank2.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank3.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank4.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank5.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/responses/rank6.jsonl +3 -0
- evaluations/base_pass8_16k_seed42/summary.json +71 -0
- evaluations/base_pass8_16k_seed42/validation.json +111 -0
- evaluations/base_pass8_16k_seed42/workers/rank0.complete.json +5 -0
- evaluations/base_pass8_16k_seed42/workers/rank0.exit_code +1 -0
- evaluations/base_pass8_16k_seed42/workers/rank0.json +40 -0
- evaluations/base_pass8_16k_seed42/workers/rank1.complete.json +5 -0
- evaluations/base_pass8_16k_seed42/workers/rank1.exit_code +1 -0
- evaluations/base_pass8_16k_seed42/workers/rank1.json +40 -0
- evaluations/base_pass8_16k_seed42/workers/rank2.complete.json +5 -0
- evaluations/base_pass8_16k_seed42/workers/rank2.exit_code +1 -0
- evaluations/base_pass8_16k_seed42/workers/rank2.json +40 -0
- evaluations/base_pass8_16k_seed42/workers/rank3.complete.json +5 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,24 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
evaluations/base_pass8_16k_seed42/responses/rank0.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
evaluations/base_pass8_16k_seed42/responses/rank1.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
evaluations/base_pass8_16k_seed42/responses/rank2.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 64 |
+
evaluations/base_pass8_16k_seed42/responses/rank3.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 65 |
+
evaluations/base_pass8_16k_seed42/responses/rank4.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 66 |
+
evaluations/base_pass8_16k_seed42/responses/rank5.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 67 |
+
evaluations/base_pass8_16k_seed42/responses/rank6.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 68 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank0.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 69 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank1.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 70 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank2.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 71 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank3.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 72 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank4.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 73 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank5.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 74 |
+
evaluations/grpo_step140_pass8_16k_seed42/responses/rank6.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 75 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank0.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 76 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank1.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 77 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank2.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 78 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank3.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 79 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank4.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 80 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank5.jsonl filter=lfs diff=lfs merge=lfs -text
|
| 81 |
+
evaluations/rltt_step140_pass8_16k_seed42/responses/rank6.jsonl filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
task_categories:
|
| 3 |
+
- text-generation
|
| 4 |
+
tags:
|
| 5 |
+
- evaluation
|
| 6 |
+
- ouro
|
| 7 |
+
- grpo
|
| 8 |
+
- rltt
|
| 9 |
+
- pass-at-8
|
| 10 |
+
configs:
|
| 11 |
+
- config_name: base
|
| 12 |
+
data_files:
|
| 13 |
+
- split: test
|
| 14 |
+
path: evaluations/base_pass8_16k_seed42/responses/*.jsonl
|
| 15 |
+
- config_name: grpo_step140
|
| 16 |
+
data_files:
|
| 17 |
+
- split: test
|
| 18 |
+
path: evaluations/grpo_step140_pass8_16k_seed42/responses/*.jsonl
|
| 19 |
+
- config_name: rltt_step140
|
| 20 |
+
data_files:
|
| 21 |
+
- split: test
|
| 22 |
+
path: evaluations/rltt_step140_pass8_16k_seed42/responses/*.jsonl
|
| 23 |
+
---
|
| 24 |
+
# Ouro-1.4B Thinking: completed base, GRPO and RLTT evaluation results
|
| 25 |
+
|
| 26 |
+
Public evaluation artifacts: frozen benchmark questions/prompts and source revisions,
|
| 27 |
+
all generated responses and token IDs, predictions/correctness, per-request seeds,
|
| 28 |
+
native grader and sampling code, final summaries, worker exit evidence and logs.
|
| 29 |
+
The checkpoint model repositories retain their existing visibility.
|
| 30 |
+
|
| 31 |
+
pass@8 is the fraction of questions with at least one correct answer in eight samples.
|
| 32 |
+
Sampling: four recurrent loops, BF16 inference, maximum 16384 new tokens,
|
| 33 |
+
temperature 0.6, top_p 0.95, top_k disabled, engine seed 42.
|
| 34 |
+
Request seeds are SHA256(protocol|uid|replicate|42)[:16] & 0x7fffffff.
|
| 35 |
+
The protocol is looprl-math-pass8-16k-v1. Last complete boxed answer is graded
|
| 36 |
+
by exact normalization or math-verify 0.8.0; the native grader is included.
|
| 37 |
+
|
| 38 |
+
Base and GRPO have 1637 questions and 13096 responses each across nine suites.
|
| 39 |
+
RLTT has 962 questions and 7696 responses across the eight requested suites.
|
| 40 |
+
RLTT was not evaluated on OlympiadBench. All committed records were checked
|
| 41 |
+
against frozen data, request seeds, unique replicate IDs 0-7 and native summaries.
|
| 42 |
+
All evaluation workers and coordinators exited 0.
|
| 43 |
+
|
| 44 |
+
Base resumed 6947 previously committed samples. Its earlier execution used
|
| 45 |
+
concurrency8/memory0.60/token-batch4096; the resumed execution used
|
| 46 |
+
concurrency16/memory0.85/token-batch32768. Both original protocols are included.
|
| 47 |
+
The original RLTT training driver has no recorded native exit code after platform
|
| 48 |
+
termination. Evaluation was explicitly authorized after full checkpoint and exact
|
| 49 |
+
CPU export verification. This does not change its unknown training exit code.
|
| 50 |
+
|
| 51 |
+
## Checkpoint repositories
|
| 52 |
+
|
| 53 |
+
- [grpo_step140](https://huggingface.co/icemoon28/Ouro-1.4B-Thinking-GRPO-step140/tree/70653159ca24c66719eae5e4edf250bf1cf9b4e4)
|
| 54 |
+
- [rltt_step140](https://huggingface.co/icemoon28/Ouro-1.4B-Thinking-RLTT-step140/tree/4dc17d193c1789a25f333c16b029af0d77a7b2d7)
|
| 55 |
+
|
| 56 |
+
## Final pass@8
|
| 57 |
+
|
| 58 |
+
| Dataset | Base | GRPO step140 | RLTT step140 |
|
| 59 |
+
|---|---:|---:|---:|
|
| 60 |
+
| aime24 | 63.33% (19/30) | 73.33% (22/30) | 76.67% (23/30) |
|
| 61 |
+
| aime25 | 60.00% (18/30) | 73.33% (22/30) | 70.00% (21/30) |
|
| 62 |
+
| aime26 | 70.00% (21/30) | 70.00% (21/30) | 70.00% (21/30) |
|
| 63 |
+
| math500 | 97.40% (487/500) | 97.80% (489/500) | 98.00% (490/500) |
|
| 64 |
+
| amc23 | 92.50% (37/40) | 95.00% (38/40) | 95.00% (38/40) |
|
| 65 |
+
| olympiadbench | 74.07% (500/675) | 76.30% (515/675) | Not evaluated |
|
| 66 |
+
| minerva_math | 54.41% (148/272) | 55.88% (152/272) | 54.04% (147/272) |
|
| 67 |
+
| hmmt_feb25 | 40.00% (12/30) | 46.67% (14/30) | 36.67% (11/30) |
|
| 68 |
+
| hmmt_nov25 | 56.67% (17/30) | 66.67% (20/30) | 60.00% (18/30) |
|
| 69 |
+
|
| 70 |
+
## Artifact directories
|
| 71 |
+
|
| 72 |
+
- [base](evaluations/base_pass8_16k_seed42): summary.json, validation.json, responses/, data/, code/, logs/ and worker evidence.
|
| 73 |
+
- [grpo_step140](evaluations/grpo_step140_pass8_16k_seed42): summary.json, validation.json, responses/, data/, code/, logs/ and worker evidence.
|
| 74 |
+
- [rltt_step140](evaluations/rltt_step140_pass8_16k_seed42): summary.json, validation.json, responses/, data/, code/, logs/ and worker evidence.
|
| 75 |
+
|
| 76 |
+
comparison.json contains exact metrics and completion checks. release-manifest.json and SHA256SUMS bind the published files. No model weights, training data, training audit traces, runtime caches or credentials are included.
|
SHA256SUMS
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0fa7f591f44d475c44391227dc990a169ce26413f8e0cd40cfd5ee99539013df README.md
|
| 2 |
+
bb5226246433dfc0cf6831c0a02986ec8f123f43b048bb7becdf3c288290c5f3 comparison.json
|
| 3 |
+
617694922ad4be85938cd17509a6f839e81d702d6d228dc4a495d88f2ab62dd0 evaluations/base_pass8_16k_seed42/code/rltt_baseline/eval_math_suite.py
|
| 4 |
+
6b477741650c722c1b9a38b538ae9b97291a98d0eb2a5fbcf9048765c76feef4 evaluations/base_pass8_16k_seed42/code/scripts/resume_math_eval.py
|
| 5 |
+
c2b6bd715b2a59aed546e2189edd9731ac065208853e63644da7f20910c169e1 evaluations/base_pass8_16k_seed42/code/scripts/run_math_eval_suite.py
|
| 6 |
+
657e9b96e78675c49d91c68d8055a6b2b421a6e5cf3b2047e8e17cebd0495f23 evaluations/base_pass8_16k_seed42/code/tests/test_math_eval_suite.py
|
| 7 |
+
65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26 evaluations/base_pass8_16k_seed42/data/questions.jsonl
|
| 8 |
+
ba7591920cd1a1f0a604eb19d41129fdeb14131454b7506d919f3f3294aded3a evaluations/base_pass8_16k_seed42/data/sources.json
|
| 9 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/eval_exit_code
|
| 10 |
+
0e8e6a6b4ccb59f2176b7f9f7a6fb3412501dfbd6592a562763b63c859d585b9 evaluations/base_pass8_16k_seed42/logs/coordinator.log
|
| 11 |
+
75b6fe15473e2eed0b6cb62a3e66a80436bc34a3e67bb1e1605000eccd08333b evaluations/base_pass8_16k_seed42/logs/rank0.log
|
| 12 |
+
12966ac38dfe6c8f09cd74dac6fbf635fc3bfdf9e4e3bfc79f346a6523042503 evaluations/base_pass8_16k_seed42/logs/rank1.log
|
| 13 |
+
e5b668f74e158b187eb8a17abe268bb73ee90ef8839db8ae8209d89be8d62c88 evaluations/base_pass8_16k_seed42/logs/rank2.log
|
| 14 |
+
27b57f708a67ca015b88900241b3b8c4dc05645628eacd469b09ae0ec2492a89 evaluations/base_pass8_16k_seed42/logs/rank3.log
|
| 15 |
+
449b7abcb80db8ae7940b33376eef4b6f752f81d395d1b5a876be906e1b5fcdc evaluations/base_pass8_16k_seed42/logs/rank4.log
|
| 16 |
+
97e4939f9c1cbf500520d1d815abd0dae571a86668da48f4583b65be5976ec10 evaluations/base_pass8_16k_seed42/logs/rank5.log
|
| 17 |
+
371ab78e9c099f0beaf44017dac3e6c37bde01e93ca869e680c5683013e0a55d evaluations/base_pass8_16k_seed42/logs/rank6.log
|
| 18 |
+
87e8d7cac5e75cc8adf200956609b64aa6b3d46e1035f2221f7232ebbe5c3d22 evaluations/base_pass8_16k_seed42/preflight/code_hashes.json
|
| 19 |
+
d8c29393425d95074bcd5b04c41286798c99aa744d89f0df9265fbd3176abd8a evaluations/base_pass8_16k_seed42/preflight/engine_rank0.json
|
| 20 |
+
454f2cafc4aef26e3fa3925f07f3bf57dd4f2c5317be12e53a4fb1243157172a evaluations/base_pass8_16k_seed42/preflight/engine_rank1.json
|
| 21 |
+
206b3828414de911df4e4bced60cc9a1e4d101e32907164864cc1783754c7add evaluations/base_pass8_16k_seed42/preflight/engine_rank2.json
|
| 22 |
+
032a540268a8bdcef7c0fe820c42e80f287e13837ebca8a8d3fac613ac04d787 evaluations/base_pass8_16k_seed42/preflight/engine_rank3.json
|
| 23 |
+
31abd9bf507ba9308d2357ea36ee9e7a3ec5a7dc0de9e21f4d6e82a94f16b049 evaluations/base_pass8_16k_seed42/preflight/engine_rank4.json
|
| 24 |
+
136d4bbedb8e51bb8948c0472ba9665a0d61eedbccfb577eae0e9be5a91d81e4 evaluations/base_pass8_16k_seed42/preflight/engine_rank5.json
|
| 25 |
+
0feab20708a9807c8c7a8356d93f52ff2b84270009eca834c6405b1d10b009b0 evaluations/base_pass8_16k_seed42/preflight/engine_rank6.json
|
| 26 |
+
a6774e9d8c0eafdd1323ed6bb4c92b2c2ce8a9b90a3951ef298dead233171151 evaluations/base_pass8_16k_seed42/preflight/launch_receipt.json
|
| 27 |
+
ec40b34c6bed8fa17311d68f1d2dc22d60d43d661c550b44db3683fb184bd6d4 evaluations/base_pass8_16k_seed42/preflight/resume.json
|
| 28 |
+
253ba69f5896ec554c0fb9ad0762e2ab18df19299c0fc2615d8989b7508fa285 evaluations/base_pass8_16k_seed42/protocol.json
|
| 29 |
+
cb56846c4b8ada93834abc9a0bada8bfa92be1187656e4ee96d151936d6386a1 evaluations/base_pass8_16k_seed42/provenance/parent_protocol.json
|
| 30 |
+
b474e6b519151dfee60ae1ed25cc884e73132a0f40b85d7648568fc5cc120880 evaluations/base_pass8_16k_seed42/responses/rank0.jsonl
|
| 31 |
+
b74eb39bc13e9863c5e00130ab8c1859f5530c6bb487ab0f58cf88143bcc0d60 evaluations/base_pass8_16k_seed42/responses/rank1.jsonl
|
| 32 |
+
d6ba8f53116beca816cca835c8107e2e5a1734a7f2188fd1ba91a8cb733d7683 evaluations/base_pass8_16k_seed42/responses/rank2.jsonl
|
| 33 |
+
31bc6e2a07cdaf5725d4a6b404b905ee55cb75b6abc78d388107826526ee7ec3 evaluations/base_pass8_16k_seed42/responses/rank3.jsonl
|
| 34 |
+
9c276b840782c58b15b43491fae031f789d255fc7156739e836db7dd979a6de0 evaluations/base_pass8_16k_seed42/responses/rank4.jsonl
|
| 35 |
+
ffe3fb47b2c7b3baeae904a432c19ccea5b3c26939578e78197636f89ed891a6 evaluations/base_pass8_16k_seed42/responses/rank5.jsonl
|
| 36 |
+
44bc336b34dae4d1352a74cfb331f4a6edf204517247880a1e4d212097dad2b8 evaluations/base_pass8_16k_seed42/responses/rank6.jsonl
|
| 37 |
+
47713208dee8cc2c50a12fb081a3ac2f08f24b09caf6247221e53fd247dfccf2 evaluations/base_pass8_16k_seed42/summary.json
|
| 38 |
+
3b6d93a6ef28a3694834dfff505ca3078b9d53398341bcf92d4ec4bb0127261b evaluations/base_pass8_16k_seed42/validation.json
|
| 39 |
+
abbf8ccea35cb99b2315969dbf6c2d52b5c400f6dd50b3e14c4a0d5440eadd76 evaluations/base_pass8_16k_seed42/workers/rank0.complete.json
|
| 40 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank0.exit_code
|
| 41 |
+
c1baabf426e4b4b4dd477acb0f33a074dbaa153863850c4fc8b5ee37a95bbc26 evaluations/base_pass8_16k_seed42/workers/rank0.json
|
| 42 |
+
2e7b532b1c7c6a1590462da03e2bb1f9ee5ef9aa5ef198a9f11291e16e2a4d99 evaluations/base_pass8_16k_seed42/workers/rank1.complete.json
|
| 43 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank1.exit_code
|
| 44 |
+
24eaed13970d18034f13782d7183861284ae45d0b88a85be355163df33530efa evaluations/base_pass8_16k_seed42/workers/rank1.json
|
| 45 |
+
167d9272ddd764cacef9a74b851bea02eae3cc637671982e059cd2ba184832c5 evaluations/base_pass8_16k_seed42/workers/rank2.complete.json
|
| 46 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank2.exit_code
|
| 47 |
+
f345b8b213fa4bd9dc5926f7a8c13cd033e14fdb6426d7385ed95694162ad0b0 evaluations/base_pass8_16k_seed42/workers/rank2.json
|
| 48 |
+
a4760b2626b213449a7e69730ab7f95eb460de12999f5c9ba5062ccd2620c817 evaluations/base_pass8_16k_seed42/workers/rank3.complete.json
|
| 49 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank3.exit_code
|
| 50 |
+
f12e05b818c648fd6a05e37d136ecaabd95c4f087c8e64c4b01731a8efcf2010 evaluations/base_pass8_16k_seed42/workers/rank3.json
|
| 51 |
+
367fb9352558e797b12f2a29df802a8b274ae1a1d9fc4093649348f900cfd050 evaluations/base_pass8_16k_seed42/workers/rank4.complete.json
|
| 52 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank4.exit_code
|
| 53 |
+
29a8760014812079fa30c532ee971d06fd860dd27433222ce48769c9f26f3639 evaluations/base_pass8_16k_seed42/workers/rank4.json
|
| 54 |
+
d68ea2710c25e2fcfc763fdd8f8c5ea7e3c8f6e10c012f3c1946c75e5d3775bc evaluations/base_pass8_16k_seed42/workers/rank5.complete.json
|
| 55 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank5.exit_code
|
| 56 |
+
c4b0ca5ff234ec0293a097bd2c9ad091c698cf0a03b17dc8147c05373ef1f128 evaluations/base_pass8_16k_seed42/workers/rank5.json
|
| 57 |
+
803cff32f36c6e1c8ae5850ad77c73fa066d12a1c50e2bb217b02e6cb763d3d5 evaluations/base_pass8_16k_seed42/workers/rank6.complete.json
|
| 58 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/base_pass8_16k_seed42/workers/rank6.exit_code
|
| 59 |
+
6b97a9212558e3f8b2a82f45fca4b4462fd20e45b9018ace87dd550b5b7ca75c evaluations/base_pass8_16k_seed42/workers/rank6.json
|
| 60 |
+
617694922ad4be85938cd17509a6f839e81d702d6d228dc4a495d88f2ab62dd0 evaluations/grpo_step140_pass8_16k_seed42/code/rltt_baseline/eval_math_suite.py
|
| 61 |
+
6b477741650c722c1b9a38b538ae9b97291a98d0eb2a5fbcf9048765c76feef4 evaluations/grpo_step140_pass8_16k_seed42/code/scripts/resume_math_eval.py
|
| 62 |
+
c2b6bd715b2a59aed546e2189edd9731ac065208853e63644da7f20910c169e1 evaluations/grpo_step140_pass8_16k_seed42/code/scripts/run_math_eval_suite.py
|
| 63 |
+
657e9b96e78675c49d91c68d8055a6b2b421a6e5cf3b2047e8e17cebd0495f23 evaluations/grpo_step140_pass8_16k_seed42/code/tests/test_math_eval_suite.py
|
| 64 |
+
65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26 evaluations/grpo_step140_pass8_16k_seed42/data/questions.jsonl
|
| 65 |
+
ba7591920cd1a1f0a604eb19d41129fdeb14131454b7506d919f3f3294aded3a evaluations/grpo_step140_pass8_16k_seed42/data/sources.json
|
| 66 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/eval_exit_code
|
| 67 |
+
44256247ec9d53f38f14a4cc84c2994bd4c17c6fa9dded35f449bd977700fffa evaluations/grpo_step140_pass8_16k_seed42/logs/coordinator.log
|
| 68 |
+
4b3d0d0acd609a21ed2cb604f7606764c323a60bf4e231c1274e9a6e646300a7 evaluations/grpo_step140_pass8_16k_seed42/logs/rank0.log
|
| 69 |
+
fb358faa8090d8dfc6a3b69b71875deae1bbb43ffacc9fd9c586991d9081277e evaluations/grpo_step140_pass8_16k_seed42/logs/rank1.log
|
| 70 |
+
7a4c3422019b8a91b2ea0221140a8b990e02c41063a5668cea8cd8466fd5c1da evaluations/grpo_step140_pass8_16k_seed42/logs/rank2.log
|
| 71 |
+
2e2b26428ad201d355e148b2c1cacd8ea46081ed29dcc83efd391f03c72f2776 evaluations/grpo_step140_pass8_16k_seed42/logs/rank3.log
|
| 72 |
+
f603547c441219c2b820c9f7c75f5804f93a6fbd04b21d425183e22e224f47cc evaluations/grpo_step140_pass8_16k_seed42/logs/rank4.log
|
| 73 |
+
ef53329119ad913c02c2714e811adc6bee7701d9255be89a921ac2c7ea805abe evaluations/grpo_step140_pass8_16k_seed42/logs/rank5.log
|
| 74 |
+
cac058fdfa573d7c1d0b3aa010745366b9590173ed769e81174aee5e72f3f26d evaluations/grpo_step140_pass8_16k_seed42/logs/rank6.log
|
| 75 |
+
31a1380c1d85c33fe6520e1d335cdeeb95ab14ac70bcce45bc519d4451539eba evaluations/grpo_step140_pass8_16k_seed42/preflight/base_result_validation.json
|
| 76 |
+
4fd786f14a906720354532e7b18b01dc10f820393d94f04bf77f3bd427a1034a evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank0.json
|
| 77 |
+
e0d9d3de9a2bd1190d5ace38650e766d1532dc58d61d576e5feddc78caf97dd0 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank1.json
|
| 78 |
+
20a9a459cf227c5c6acb670247fe7eb0950dce1cd6b521510a01a478cf6261e3 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank2.json
|
| 79 |
+
f86e0f019c6fc0d84b2365be1b6fad90cccec80d56e4e8c036251d908d07e816 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank3.json
|
| 80 |
+
f64f6fa403365439d7ac9af1d9dfd258011bfcf18b9a0e65f22e31c51fc14402 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank4.json
|
| 81 |
+
b6481c2797df3b288cb7b08b9b50fd1430e39965c134cd3cfa4108c84e701060 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank5.json
|
| 82 |
+
dbe54e668b9c85a844aa66fab78c2d5abaed7b7c40e9b5f0b1b2e311a97b8115 evaluations/grpo_step140_pass8_16k_seed42/preflight/engine_rank6.json
|
| 83 |
+
688dc49f942b18e5bbc8e9e67c12b5912838c213f8fc6d974cac7f80fd396cae evaluations/grpo_step140_pass8_16k_seed42/preflight/launch_receipt.json
|
| 84 |
+
a3520be86f20d53dccb944b4816cd81bb5923e9fcb9ae092cc3234656e6043b0 evaluations/grpo_step140_pass8_16k_seed42/preflight/preparation.json
|
| 85 |
+
5470a0f7b2429408f74a879c494baddab26319f6e500d318bf5b6c4492a6f017 evaluations/grpo_step140_pass8_16k_seed42/protocol.json
|
| 86 |
+
6c6a2d91c3e566098e8f16550c4b165e6e1f7de405e53f46bc1c9b080f29ada0 evaluations/grpo_step140_pass8_16k_seed42/responses/rank0.jsonl
|
| 87 |
+
ec2f9733cd9fbdc0ee8068d05ed58b1ba6bde7836ec1fbb498982a868a25cbbf evaluations/grpo_step140_pass8_16k_seed42/responses/rank1.jsonl
|
| 88 |
+
aeb857c63c87ab84a555936b092e86cab3427aa634c249601ba6c3a123083f07 evaluations/grpo_step140_pass8_16k_seed42/responses/rank2.jsonl
|
| 89 |
+
69dd24c09075fb67afdfc092cf43e263939d7267a15a6d7a048210a6f29a6f23 evaluations/grpo_step140_pass8_16k_seed42/responses/rank3.jsonl
|
| 90 |
+
5aa2169552280634a4fb8fd0535d1926f572228dfe389f9535cc20ba2cf6a450 evaluations/grpo_step140_pass8_16k_seed42/responses/rank4.jsonl
|
| 91 |
+
d2906c87bb62793986fee71f8a52bcd5e9d55f4413c543090c10590417fffe4b evaluations/grpo_step140_pass8_16k_seed42/responses/rank5.jsonl
|
| 92 |
+
34cf68769f3108644eeb8776f46cf6e2bc13e372207a8c1452ef7af3c5c1b0fd evaluations/grpo_step140_pass8_16k_seed42/responses/rank6.jsonl
|
| 93 |
+
dad76a912991f0086fd888d61c4ede5c5a7d57e5aa2d5c17d8a6cb0ff39ec312 evaluations/grpo_step140_pass8_16k_seed42/summary.json
|
| 94 |
+
9786e0b8e19a9afe32d926e7c780b0986dae41b8ffbafcc241005f33b7efc2ec evaluations/grpo_step140_pass8_16k_seed42/validation.json
|
| 95 |
+
aae2e53a25a310d7c183817f041244215c44fe469bb2a85308dc383d179325e1 evaluations/grpo_step140_pass8_16k_seed42/workers/rank0.complete.json
|
| 96 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank0.exit_code
|
| 97 |
+
2f1591a443351edb8fe4908cd2af29c0e48718f0668b041dc42143c003a30535 evaluations/grpo_step140_pass8_16k_seed42/workers/rank0.json
|
| 98 |
+
f65702a7e3a14d768bf84b94d1fbe1643fa5baf514d8d6f6f45d647b92a4c76a evaluations/grpo_step140_pass8_16k_seed42/workers/rank1.complete.json
|
| 99 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank1.exit_code
|
| 100 |
+
27143b196e698d3749f36b1b62f1ac9cd300537bbc2bade8193464256cf31408 evaluations/grpo_step140_pass8_16k_seed42/workers/rank1.json
|
| 101 |
+
a850ad02609e48a0e210ecb42d55bb2a282b1bd839eb4f727b1bdb6c1aa9bb3b evaluations/grpo_step140_pass8_16k_seed42/workers/rank2.complete.json
|
| 102 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank2.exit_code
|
| 103 |
+
fb839221aa67a23e7a3ad8d016d334ce4ea6846af290b9699883df71b869f3f3 evaluations/grpo_step140_pass8_16k_seed42/workers/rank2.json
|
| 104 |
+
3803ff0f77af8d3a844fb4b9a52ff79c410a801cf7dca4a7b1a56b8bafaa732c evaluations/grpo_step140_pass8_16k_seed42/workers/rank3.complete.json
|
| 105 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank3.exit_code
|
| 106 |
+
cf3cb6b1cdb75aafd46dfd089dcce3e98445f6e22f0b47c0f7aae178c8dbbcdf evaluations/grpo_step140_pass8_16k_seed42/workers/rank3.json
|
| 107 |
+
e3e3c780c8f308744027cef7480c47a09cae8aa541815afc4ddf422464929e0d evaluations/grpo_step140_pass8_16k_seed42/workers/rank4.complete.json
|
| 108 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank4.exit_code
|
| 109 |
+
39f280ee8c9c6f6e6d7ed1f68a7985c014e69c890fd4fa3774800bf2051fddcf evaluations/grpo_step140_pass8_16k_seed42/workers/rank4.json
|
| 110 |
+
f3462ad828f34bd3c9fccbc3c9c60f6a0f66b11bdc7dedb4d693763971723f56 evaluations/grpo_step140_pass8_16k_seed42/workers/rank5.complete.json
|
| 111 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank5.exit_code
|
| 112 |
+
f5ae66e319eee136839cc3452e29790a1bea3489452225871fbca21c7a524736 evaluations/grpo_step140_pass8_16k_seed42/workers/rank5.json
|
| 113 |
+
5c4c822a64b2cf5c4f22648751c99eb64a265ecc6ee7a7048edfbb0d7562e686 evaluations/grpo_step140_pass8_16k_seed42/workers/rank6.complete.json
|
| 114 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/grpo_step140_pass8_16k_seed42/workers/rank6.exit_code
|
| 115 |
+
d529b17c3c627beb692b2c74e5ad7a5e0db5850fc3febd17002977d8a42bf651 evaluations/grpo_step140_pass8_16k_seed42/workers/rank6.json
|
| 116 |
+
af873c9be6497ed252ff9e6eea6cc6885f57f2076b7486013fb9daf05708acb3 evaluations/rltt_step140_pass8_16k_seed42/code/rltt_baseline/eval_math_suite.py
|
| 117 |
+
c2b6bd715b2a59aed546e2189edd9731ac065208853e63644da7f20910c169e1 evaluations/rltt_step140_pass8_16k_seed42/code/run_math_eval_suite.py
|
| 118 |
+
80e4d7c756243aed7eca211cace5174ac2497663c5202f2de4b9d8de15e58932 evaluations/rltt_step140_pass8_16k_seed42/data/questions.jsonl
|
| 119 |
+
769d5b5a7cee58be00d6c1e39f37bb89dbebf7c321dc51359a0e2729a7ccaaba evaluations/rltt_step140_pass8_16k_seed42/data/sources.json
|
| 120 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/eval_exit_code
|
| 121 |
+
a5f32743351d5a282185e5fb00d962016564e59c498136295abbcaf400d73354 evaluations/rltt_step140_pass8_16k_seed42/logs/coordinator.log
|
| 122 |
+
f5122a2f33ba097044c36a3b45fee09db6c3b79298ada4a1900ef3e0c6f4c183 evaluations/rltt_step140_pass8_16k_seed42/logs/rank0.log
|
| 123 |
+
e34920984b16cc350388c0f72bf56937e5586816d15c5f8a155df78e004e484c evaluations/rltt_step140_pass8_16k_seed42/logs/rank1.log
|
| 124 |
+
7d306cc01053bfe4ef8c0c2564eeb3da0f20e6ca8f072bcb6a92c55fad0db7eb evaluations/rltt_step140_pass8_16k_seed42/logs/rank2.log
|
| 125 |
+
11c39e9ab70911af48eaef9535d2f8b07f6f2e5284488d05f78c821cfacc0f16 evaluations/rltt_step140_pass8_16k_seed42/logs/rank3.log
|
| 126 |
+
7b2e6e4467111c67d3308ceefcc75930becb76ec1994be6a3e0b9b103492fbc5 evaluations/rltt_step140_pass8_16k_seed42/logs/rank4.log
|
| 127 |
+
3626fcb08ef53dc9c15e86921cfdd73f8cc45b438f44e88cdeb0659c54ebaae7 evaluations/rltt_step140_pass8_16k_seed42/logs/rank5.log
|
| 128 |
+
12a288536dc65368f5e8ef38adaaaf9009faec0355a3a58eb7cb17266ab4c406 evaluations/rltt_step140_pass8_16k_seed42/logs/rank6.log
|
| 129 |
+
c3e236667a0df8d4699a04f0b86e38ffe5eb27b2cd19a79b7850c99edf87caef evaluations/rltt_step140_pass8_16k_seed42/preflight/data.json
|
| 130 |
+
06f90ac84e3486d1ae8f0432bf45443c1155c2ed228de075796de74bf7e4a7b7 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank0.json
|
| 131 |
+
1fd7c64308607082f2be36d3d3d3ef5f175f49803b68c986a057ed3ff1471340 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank1.json
|
| 132 |
+
b3f4f9700c44e385b7b13de8a0c0a161d6b1585dae9e1c1885db910bfdd4bbb7 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank2.json
|
| 133 |
+
9b2b8319bfd7599ea129272fe384b59c4cc5e879e3d54e7f6a64e55ef3952975 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank3.json
|
| 134 |
+
70d7a099b199177c4161630d89497593f797892b99d04d812d947626ce3ec712 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank4.json
|
| 135 |
+
c57e3296c4a303128751d30f2231d3257111f43313fc682f4ab67b84fab33022 evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank5.json
|
| 136 |
+
4ee5d5a559e67736679ab2138aab0523bcc996c52aa73d2f6fbc106165625c0a evaluations/rltt_step140_pass8_16k_seed42/preflight/engine_rank6.json
|
| 137 |
+
36486a9cc9e135b0005416de0cc4477feb0c9dee760819110a0a129fe6da4146 evaluations/rltt_step140_pass8_16k_seed42/preflight/launch_receipt.json
|
| 138 |
+
adefd340ce3e4c743cbbc2e4926f634e162e3a9d20cadd191c92ec4eb6150055 evaluations/rltt_step140_pass8_16k_seed42/preflight/model.json
|
| 139 |
+
050836d5b602f7f65a2d37e806e845c13ac9413a3506fe625e9aac8e0101b751 evaluations/rltt_step140_pass8_16k_seed42/protocol.json
|
| 140 |
+
34ace94e449249d71667472e0857c690e77f990e4f10805720f82bd0fffdca05 evaluations/rltt_step140_pass8_16k_seed42/responses/rank0.jsonl
|
| 141 |
+
a0a8730c13611d9e57c5d4761f145460b4ae5571cb1c2e9c62024d86a19b1fb8 evaluations/rltt_step140_pass8_16k_seed42/responses/rank1.jsonl
|
| 142 |
+
5d862b84ddc88cd0d8a86e12080f5f91e653761d6e966a0749bf5ce0d6fe5d71 evaluations/rltt_step140_pass8_16k_seed42/responses/rank2.jsonl
|
| 143 |
+
a85a6fc0276c48e41304e953f596170491a7455cf6c0f2a2034d6f3c1d9c8f96 evaluations/rltt_step140_pass8_16k_seed42/responses/rank3.jsonl
|
| 144 |
+
40b35bd92c6207533daf337b0c6ccc76c047fe65e57590cfd993c6624b1b4d65 evaluations/rltt_step140_pass8_16k_seed42/responses/rank4.jsonl
|
| 145 |
+
6cb1f8eb5168d946be7c156eceac18678b18ffe5d5d4c071c2646c6e26776519 evaluations/rltt_step140_pass8_16k_seed42/responses/rank5.jsonl
|
| 146 |
+
2e1f42f1b4dc1d38888f53a4249fe00884c8f8e9bcacfd48c5e250a795ec2137 evaluations/rltt_step140_pass8_16k_seed42/responses/rank6.jsonl
|
| 147 |
+
b6a977cd2f322f48c18d21417e040e9fc404d91eae8c2a2b004e758ea5b2ce0c evaluations/rltt_step140_pass8_16k_seed42/summary.json
|
| 148 |
+
9c66d7322a7ab52222b5f399bc1128e793ecc0d068d25b789518abc962f1bcfa evaluations/rltt_step140_pass8_16k_seed42/validation.json
|
| 149 |
+
3521ba6219f33962badda22cb45a9555a74c3a7f91fcf4e8e1e95fdbdfdebf71 evaluations/rltt_step140_pass8_16k_seed42/workers/rank0.complete.json
|
| 150 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank0.exit_code
|
| 151 |
+
bc4464113bc2e3ec18d56f7597eb4bbe3e60ab2aa6a26835ad3163dfdd243735 evaluations/rltt_step140_pass8_16k_seed42/workers/rank0.json
|
| 152 |
+
6541e8415e52b204941c9ebdfc1f852ba5a2e462840ca12730a7525f7f90469e evaluations/rltt_step140_pass8_16k_seed42/workers/rank1.complete.json
|
| 153 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank1.exit_code
|
| 154 |
+
3b5e3447eed2530ecfe14d09c715841ac72866a911f2d930ef3d2579f7986ba6 evaluations/rltt_step140_pass8_16k_seed42/workers/rank1.json
|
| 155 |
+
4c45f161f10000a358dfa95ccfc74a78d44ff96470169f03b548163bdf942d65 evaluations/rltt_step140_pass8_16k_seed42/workers/rank2.complete.json
|
| 156 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank2.exit_code
|
| 157 |
+
390d74fe25c106bc54697bdb076e10d1af6991af62770bc2e6cbbbe25e35d917 evaluations/rltt_step140_pass8_16k_seed42/workers/rank2.json
|
| 158 |
+
f140b6801a1417d014117444887f7d7f53d753ae08697a0726b87c6d3ae3da43 evaluations/rltt_step140_pass8_16k_seed42/workers/rank3.complete.json
|
| 159 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank3.exit_code
|
| 160 |
+
34f8fd97acbc563c8f17357475c74fae7fa9d23890aabe87cff3bccc6ebdb8ae evaluations/rltt_step140_pass8_16k_seed42/workers/rank3.json
|
| 161 |
+
e8d2b5fc0b372b6301db084f2b492c1d8b5a38568019a76a839a10ac8b59da9a evaluations/rltt_step140_pass8_16k_seed42/workers/rank4.complete.json
|
| 162 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank4.exit_code
|
| 163 |
+
f37adffcd435847928202908d546abe64c66e4f5b65f3e95775fa137a612ae39 evaluations/rltt_step140_pass8_16k_seed42/workers/rank4.json
|
| 164 |
+
782d24be13b7214ccc222544444096d0d1f5c26f1cf7ed793d288ed0f3092ca8 evaluations/rltt_step140_pass8_16k_seed42/workers/rank5.complete.json
|
| 165 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank5.exit_code
|
| 166 |
+
dc1b5c58d51afdd3073ecda84f43aaa1c08698f1488238e7646b3a467742a615 evaluations/rltt_step140_pass8_16k_seed42/workers/rank5.json
|
| 167 |
+
a6fa9e819af9ceb21b55df3a7cc8307d483ce7f914f76f582f14c9e3792862ba evaluations/rltt_step140_pass8_16k_seed42/workers/rank6.complete.json
|
| 168 |
+
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evaluations/rltt_step140_pass8_16k_seed42/workers/rank6.exit_code
|
| 169 |
+
3327314e02978bfafc8c71a00359248cc0d18b4343f7ee0504c967ec4f5b9a6c evaluations/rltt_step140_pass8_16k_seed42/workers/rank6.json
|
| 170 |
+
25c1ffe9840c558e136d04210edc5525e8ae6ee1a5080b185d1e49925a0509e6 model-references.json
|
| 171 |
+
217d377a913be00487280d893d7fd5527f91128797c4371ea1b38eeaf925c091 release-manifest.json
|
comparison.json
ADDED
|
@@ -0,0 +1,342 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metric": "pass@8",
|
| 3 |
+
"definition": "Fraction of questions with at least one correct response among eight independent samples",
|
| 4 |
+
"base_seed": 42,
|
| 5 |
+
"max_new_tokens": 16384,
|
| 6 |
+
"loops": 4,
|
| 7 |
+
"dtype": "bfloat16",
|
| 8 |
+
"evaluations": {
|
| 9 |
+
"base": {
|
| 10 |
+
"status": "complete_verified",
|
| 11 |
+
"questions": 1637,
|
| 12 |
+
"responses": 13096,
|
| 13 |
+
"strict_journals": true,
|
| 14 |
+
"unique_complete_replicates": true,
|
| 15 |
+
"summary_recomputed_equal": true,
|
| 16 |
+
"worker_exit_codes": [
|
| 17 |
+
0,
|
| 18 |
+
0,
|
| 19 |
+
0,
|
| 20 |
+
0,
|
| 21 |
+
0,
|
| 22 |
+
0,
|
| 23 |
+
0
|
| 24 |
+
],
|
| 25 |
+
"coordinator_exit_code": 0,
|
| 26 |
+
"results": {
|
| 27 |
+
"aime24": {
|
| 28 |
+
"correct_questions": 19,
|
| 29 |
+
"questions": 30,
|
| 30 |
+
"pass_at_8": 0.6333333333333333
|
| 31 |
+
},
|
| 32 |
+
"aime25": {
|
| 33 |
+
"correct_questions": 18,
|
| 34 |
+
"questions": 30,
|
| 35 |
+
"pass_at_8": 0.6
|
| 36 |
+
},
|
| 37 |
+
"aime26": {
|
| 38 |
+
"correct_questions": 21,
|
| 39 |
+
"questions": 30,
|
| 40 |
+
"pass_at_8": 0.7
|
| 41 |
+
},
|
| 42 |
+
"math500": {
|
| 43 |
+
"correct_questions": 487,
|
| 44 |
+
"questions": 500,
|
| 45 |
+
"pass_at_8": 0.974
|
| 46 |
+
},
|
| 47 |
+
"amc23": {
|
| 48 |
+
"correct_questions": 37,
|
| 49 |
+
"questions": 40,
|
| 50 |
+
"pass_at_8": 0.925
|
| 51 |
+
},
|
| 52 |
+
"olympiadbench": {
|
| 53 |
+
"correct_questions": 500,
|
| 54 |
+
"questions": 675,
|
| 55 |
+
"pass_at_8": 0.7407407407407407
|
| 56 |
+
},
|
| 57 |
+
"minerva_math": {
|
| 58 |
+
"correct_questions": 148,
|
| 59 |
+
"questions": 272,
|
| 60 |
+
"pass_at_8": 0.5441176470588235
|
| 61 |
+
},
|
| 62 |
+
"hmmt_feb25": {
|
| 63 |
+
"correct_questions": 12,
|
| 64 |
+
"questions": 30,
|
| 65 |
+
"pass_at_8": 0.4
|
| 66 |
+
},
|
| 67 |
+
"hmmt_nov25": {
|
| 68 |
+
"correct_questions": 17,
|
| 69 |
+
"questions": 30,
|
| 70 |
+
"pass_at_8": 0.5666666666666667
|
| 71 |
+
}
|
| 72 |
+
},
|
| 73 |
+
"journals": {
|
| 74 |
+
"rank0.jsonl": {
|
| 75 |
+
"responses": 1872,
|
| 76 |
+
"bytes": 82830095,
|
| 77 |
+
"sha256": "b474e6b519151dfee60ae1ed25cc884e73132a0f40b85d7648568fc5cc120880",
|
| 78 |
+
"git_blob_sha1": "116337724f55a34afcb5ec87ba12c73af9111db8"
|
| 79 |
+
},
|
| 80 |
+
"rank1.jsonl": {
|
| 81 |
+
"responses": 1872,
|
| 82 |
+
"bytes": 87703153,
|
| 83 |
+
"sha256": "b74eb39bc13e9863c5e00130ab8c1859f5530c6bb487ab0f58cf88143bcc0d60",
|
| 84 |
+
"git_blob_sha1": "e4492198f4fbeffbb5792380b7c913b3f03d01cd"
|
| 85 |
+
},
|
| 86 |
+
"rank2.jsonl": {
|
| 87 |
+
"responses": 1872,
|
| 88 |
+
"bytes": 91204987,
|
| 89 |
+
"sha256": "d6ba8f53116beca816cca835c8107e2e5a1734a7f2188fd1ba91a8cb733d7683",
|
| 90 |
+
"git_blob_sha1": "d5d40a6db9dc139a6d09861d2d6be998192f84cc"
|
| 91 |
+
},
|
| 92 |
+
"rank3.jsonl": {
|
| 93 |
+
"responses": 1872,
|
| 94 |
+
"bytes": 88851916,
|
| 95 |
+
"sha256": "31bc6e2a07cdaf5725d4a6b404b905ee55cb75b6abc78d388107826526ee7ec3",
|
| 96 |
+
"git_blob_sha1": "358a8ecef0d13d8d62cf16c9e89e0bbcc815e88a"
|
| 97 |
+
},
|
| 98 |
+
"rank4.jsonl": {
|
| 99 |
+
"responses": 1872,
|
| 100 |
+
"bytes": 92933776,
|
| 101 |
+
"sha256": "9c276b840782c58b15b43491fae031f789d255fc7156739e836db7dd979a6de0",
|
| 102 |
+
"git_blob_sha1": "d3c0eb742d21a455b1a7abeb69c0801134130cbc"
|
| 103 |
+
},
|
| 104 |
+
"rank5.jsonl": {
|
| 105 |
+
"responses": 1872,
|
| 106 |
+
"bytes": 85207799,
|
| 107 |
+
"sha256": "ffe3fb47b2c7b3baeae904a432c19ccea5b3c26939578e78197636f89ed891a6",
|
| 108 |
+
"git_blob_sha1": "75a0db5c7a16a47bd4db249a57700f37f1f9415d"
|
| 109 |
+
},
|
| 110 |
+
"rank6.jsonl": {
|
| 111 |
+
"responses": 1864,
|
| 112 |
+
"bytes": 90354521,
|
| 113 |
+
"sha256": "44bc336b34dae4d1352a74cfb331f4a6edf204517247880a1e4d212097dad2b8",
|
| 114 |
+
"git_blob_sha1": "537ede571f973e2ed0be5ab6478684afcad01b57"
|
| 115 |
+
}
|
| 116 |
+
},
|
| 117 |
+
"protocol_sha256": "253ba69f5896ec554c0fb9ad0762e2ab18df19299c0fc2615d8989b7508fa285",
|
| 118 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26"
|
| 119 |
+
},
|
| 120 |
+
"grpo_step140": {
|
| 121 |
+
"status": "complete_verified",
|
| 122 |
+
"questions": 1637,
|
| 123 |
+
"responses": 13096,
|
| 124 |
+
"strict_journals": true,
|
| 125 |
+
"unique_complete_replicates": true,
|
| 126 |
+
"summary_recomputed_equal": true,
|
| 127 |
+
"worker_exit_codes": [
|
| 128 |
+
0,
|
| 129 |
+
0,
|
| 130 |
+
0,
|
| 131 |
+
0,
|
| 132 |
+
0,
|
| 133 |
+
0,
|
| 134 |
+
0
|
| 135 |
+
],
|
| 136 |
+
"coordinator_exit_code": 0,
|
| 137 |
+
"results": {
|
| 138 |
+
"aime24": {
|
| 139 |
+
"correct_questions": 22,
|
| 140 |
+
"questions": 30,
|
| 141 |
+
"pass_at_8": 0.7333333333333333
|
| 142 |
+
},
|
| 143 |
+
"aime25": {
|
| 144 |
+
"correct_questions": 22,
|
| 145 |
+
"questions": 30,
|
| 146 |
+
"pass_at_8": 0.7333333333333333
|
| 147 |
+
},
|
| 148 |
+
"aime26": {
|
| 149 |
+
"correct_questions": 21,
|
| 150 |
+
"questions": 30,
|
| 151 |
+
"pass_at_8": 0.7
|
| 152 |
+
},
|
| 153 |
+
"math500": {
|
| 154 |
+
"correct_questions": 489,
|
| 155 |
+
"questions": 500,
|
| 156 |
+
"pass_at_8": 0.978
|
| 157 |
+
},
|
| 158 |
+
"amc23": {
|
| 159 |
+
"correct_questions": 38,
|
| 160 |
+
"questions": 40,
|
| 161 |
+
"pass_at_8": 0.95
|
| 162 |
+
},
|
| 163 |
+
"olympiadbench": {
|
| 164 |
+
"correct_questions": 515,
|
| 165 |
+
"questions": 675,
|
| 166 |
+
"pass_at_8": 0.762962962962963
|
| 167 |
+
},
|
| 168 |
+
"minerva_math": {
|
| 169 |
+
"correct_questions": 152,
|
| 170 |
+
"questions": 272,
|
| 171 |
+
"pass_at_8": 0.5588235294117647
|
| 172 |
+
},
|
| 173 |
+
"hmmt_feb25": {
|
| 174 |
+
"correct_questions": 14,
|
| 175 |
+
"questions": 30,
|
| 176 |
+
"pass_at_8": 0.4666666666666667
|
| 177 |
+
},
|
| 178 |
+
"hmmt_nov25": {
|
| 179 |
+
"correct_questions": 20,
|
| 180 |
+
"questions": 30,
|
| 181 |
+
"pass_at_8": 0.6666666666666666
|
| 182 |
+
}
|
| 183 |
+
},
|
| 184 |
+
"journals": {
|
| 185 |
+
"rank0.jsonl": {
|
| 186 |
+
"responses": 1872,
|
| 187 |
+
"bytes": 79513675,
|
| 188 |
+
"sha256": "6c6a2d91c3e566098e8f16550c4b165e6e1f7de405e53f46bc1c9b080f29ada0",
|
| 189 |
+
"git_blob_sha1": "cd765a25361a89a0f67714c93db6c5561f875507"
|
| 190 |
+
},
|
| 191 |
+
"rank1.jsonl": {
|
| 192 |
+
"responses": 1872,
|
| 193 |
+
"bytes": 82259082,
|
| 194 |
+
"sha256": "ec2f9733cd9fbdc0ee8068d05ed58b1ba6bde7836ec1fbb498982a868a25cbbf",
|
| 195 |
+
"git_blob_sha1": "85f8b794d4f9feddcb62ea5184d7c393784fdcd9"
|
| 196 |
+
},
|
| 197 |
+
"rank2.jsonl": {
|
| 198 |
+
"responses": 1872,
|
| 199 |
+
"bytes": 83618172,
|
| 200 |
+
"sha256": "aeb857c63c87ab84a555936b092e86cab3427aa634c249601ba6c3a123083f07",
|
| 201 |
+
"git_blob_sha1": "b4c1bd2d65b04ad478934f3da80527c9979a1037"
|
| 202 |
+
},
|
| 203 |
+
"rank3.jsonl": {
|
| 204 |
+
"responses": 1872,
|
| 205 |
+
"bytes": 81842453,
|
| 206 |
+
"sha256": "69dd24c09075fb67afdfc092cf43e263939d7267a15a6d7a048210a6f29a6f23",
|
| 207 |
+
"git_blob_sha1": "2d7d01a8621fcf4d08f26f335a7829097a39ffad"
|
| 208 |
+
},
|
| 209 |
+
"rank4.jsonl": {
|
| 210 |
+
"responses": 1872,
|
| 211 |
+
"bytes": 87442856,
|
| 212 |
+
"sha256": "5aa2169552280634a4fb8fd0535d1926f572228dfe389f9535cc20ba2cf6a450",
|
| 213 |
+
"git_blob_sha1": "0802e23ce3ae4e13bdde1f7bdeb47f7bd8c562af"
|
| 214 |
+
},
|
| 215 |
+
"rank5.jsonl": {
|
| 216 |
+
"responses": 1872,
|
| 217 |
+
"bytes": 80133312,
|
| 218 |
+
"sha256": "d2906c87bb62793986fee71f8a52bcd5e9d55f4413c543090c10590417fffe4b",
|
| 219 |
+
"git_blob_sha1": "fb5952bf35aa6b08a177477541e16dbb766ccb41"
|
| 220 |
+
},
|
| 221 |
+
"rank6.jsonl": {
|
| 222 |
+
"responses": 1864,
|
| 223 |
+
"bytes": 84329902,
|
| 224 |
+
"sha256": "34cf68769f3108644eeb8776f46cf6e2bc13e372207a8c1452ef7af3c5c1b0fd",
|
| 225 |
+
"git_blob_sha1": "23ba2c6c9bc33de1f54d462190d36df7cb1718e0"
|
| 226 |
+
}
|
| 227 |
+
},
|
| 228 |
+
"protocol_sha256": "5470a0f7b2429408f74a879c494baddab26319f6e500d318bf5b6c4492a6f017",
|
| 229 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26"
|
| 230 |
+
},
|
| 231 |
+
"rltt_step140": {
|
| 232 |
+
"status": "complete_verified",
|
| 233 |
+
"questions": 962,
|
| 234 |
+
"responses": 7696,
|
| 235 |
+
"strict_journals": true,
|
| 236 |
+
"unique_complete_replicates": true,
|
| 237 |
+
"summary_recomputed_equal": true,
|
| 238 |
+
"worker_exit_codes": [
|
| 239 |
+
0,
|
| 240 |
+
0,
|
| 241 |
+
0,
|
| 242 |
+
0,
|
| 243 |
+
0,
|
| 244 |
+
0,
|
| 245 |
+
0
|
| 246 |
+
],
|
| 247 |
+
"coordinator_exit_code": 0,
|
| 248 |
+
"results": {
|
| 249 |
+
"aime24": {
|
| 250 |
+
"correct_questions": 23,
|
| 251 |
+
"questions": 30,
|
| 252 |
+
"pass_at_8": 0.7666666666666667
|
| 253 |
+
},
|
| 254 |
+
"aime25": {
|
| 255 |
+
"correct_questions": 21,
|
| 256 |
+
"questions": 30,
|
| 257 |
+
"pass_at_8": 0.7
|
| 258 |
+
},
|
| 259 |
+
"aime26": {
|
| 260 |
+
"correct_questions": 21,
|
| 261 |
+
"questions": 30,
|
| 262 |
+
"pass_at_8": 0.7
|
| 263 |
+
},
|
| 264 |
+
"math500": {
|
| 265 |
+
"correct_questions": 490,
|
| 266 |
+
"questions": 500,
|
| 267 |
+
"pass_at_8": 0.98
|
| 268 |
+
},
|
| 269 |
+
"amc23": {
|
| 270 |
+
"correct_questions": 38,
|
| 271 |
+
"questions": 40,
|
| 272 |
+
"pass_at_8": 0.95
|
| 273 |
+
},
|
| 274 |
+
"minerva_math": {
|
| 275 |
+
"correct_questions": 147,
|
| 276 |
+
"questions": 272,
|
| 277 |
+
"pass_at_8": 0.5404411764705882
|
| 278 |
+
},
|
| 279 |
+
"hmmt_feb25": {
|
| 280 |
+
"correct_questions": 11,
|
| 281 |
+
"questions": 30,
|
| 282 |
+
"pass_at_8": 0.36666666666666664
|
| 283 |
+
},
|
| 284 |
+
"hmmt_nov25": {
|
| 285 |
+
"correct_questions": 18,
|
| 286 |
+
"questions": 30,
|
| 287 |
+
"pass_at_8": 0.6
|
| 288 |
+
}
|
| 289 |
+
},
|
| 290 |
+
"journals": {
|
| 291 |
+
"rank0.jsonl": {
|
| 292 |
+
"responses": 1104,
|
| 293 |
+
"bytes": 40098944,
|
| 294 |
+
"sha256": "34ace94e449249d71667472e0857c690e77f990e4f10805720f82bd0fffdca05",
|
| 295 |
+
"git_blob_sha1": "cd4b763b473a5be9ec50b757aa9d0fa087ab7b69"
|
| 296 |
+
},
|
| 297 |
+
"rank1.jsonl": {
|
| 298 |
+
"responses": 1104,
|
| 299 |
+
"bytes": 43325239,
|
| 300 |
+
"sha256": "a0a8730c13611d9e57c5d4761f145460b4ae5571cb1c2e9c62024d86a19b1fb8",
|
| 301 |
+
"git_blob_sha1": "7e192d17f0854fb6ce7dd5380c6728560f1bda37"
|
| 302 |
+
},
|
| 303 |
+
"rank2.jsonl": {
|
| 304 |
+
"responses": 1104,
|
| 305 |
+
"bytes": 41917997,
|
| 306 |
+
"sha256": "5d862b84ddc88cd0d8a86e12080f5f91e653761d6e966a0749bf5ce0d6fe5d71",
|
| 307 |
+
"git_blob_sha1": "887a95b455436943ce004efae224962636790fc3"
|
| 308 |
+
},
|
| 309 |
+
"rank3.jsonl": {
|
| 310 |
+
"responses": 1096,
|
| 311 |
+
"bytes": 42039902,
|
| 312 |
+
"sha256": "a85a6fc0276c48e41304e953f596170491a7455cf6c0f2a2034d6f3c1d9c8f96",
|
| 313 |
+
"git_blob_sha1": "5e1ca29f3bd408002f1516d4cacf8e3773727d2c"
|
| 314 |
+
},
|
| 315 |
+
"rank4.jsonl": {
|
| 316 |
+
"responses": 1096,
|
| 317 |
+
"bytes": 44091376,
|
| 318 |
+
"sha256": "40b35bd92c6207533daf337b0c6ccc76c047fe65e57590cfd993c6624b1b4d65",
|
| 319 |
+
"git_blob_sha1": "879858d425f43711256a610203ec1a56965a9b07"
|
| 320 |
+
},
|
| 321 |
+
"rank5.jsonl": {
|
| 322 |
+
"responses": 1096,
|
| 323 |
+
"bytes": 39133614,
|
| 324 |
+
"sha256": "6cb1f8eb5168d946be7c156eceac18678b18ffe5d5d4c071c2646c6e26776519",
|
| 325 |
+
"git_blob_sha1": "480842493fa543bcaa109c45054d95829115ad1a"
|
| 326 |
+
},
|
| 327 |
+
"rank6.jsonl": {
|
| 328 |
+
"responses": 1096,
|
| 329 |
+
"bytes": 36385700,
|
| 330 |
+
"sha256": "2e1f42f1b4dc1d38888f53a4249fe00884c8f8e9bcacfd48c5e250a795ec2137",
|
| 331 |
+
"git_blob_sha1": "2e610453b8fc769255bcc7ec4567ca81a0a27af0"
|
| 332 |
+
}
|
| 333 |
+
},
|
| 334 |
+
"protocol_sha256": "050836d5b602f7f65a2d37e806e845c13ac9413a3506fe625e9aac8e0101b751",
|
| 335 |
+
"questions_sha256": "80e4d7c756243aed7eca211cace5174ac2497663c5202f2de4b9d8de15e58932"
|
| 336 |
+
}
|
| 337 |
+
},
|
| 338 |
+
"scope": "Base and GRPO each contain nine suites; RLTT contains the eight explicitly selected suites and no OlympiadBench.",
|
| 339 |
+
"base_runtime_note": "Base resumes 6947 committed samples from an earlier execution with different runtime batching; both protocols are included.",
|
| 340 |
+
"rltt_native_training_exit_code": null,
|
| 341 |
+
"rltt_exit_unknown_posttrain_explicitly_authorized": true
|
| 342 |
+
}
|
evaluations/base_pass8_16k_seed42/code/rltt_baseline/eval_math_suite.py
ADDED
|
@@ -0,0 +1,390 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Resumable, fixed-policy math evaluation: eight samples, 16,384 output tokens.
|
| 2 |
+
|
| 3 |
+
Use ``prepare`` on a directory containing data/sources.json, then run independent
|
| 4 |
+
``worker`` ranks on explicitly selected GPUs. No training or W&B process is used.
|
| 5 |
+
Each worker admits the frozen protocol's concurrency limit into vLLM's
|
| 6 |
+
continuous batch. Pass@8 is a sampling count, independent of concurrency.
|
| 7 |
+
"""
|
| 8 |
+
|
| 9 |
+
import argparse
|
| 10 |
+
import collections
|
| 11 |
+
import dataclasses
|
| 12 |
+
import fcntl
|
| 13 |
+
import hashlib
|
| 14 |
+
import importlib.metadata
|
| 15 |
+
import json
|
| 16 |
+
import math
|
| 17 |
+
import os
|
| 18 |
+
from pathlib import Path
|
| 19 |
+
import re
|
| 20 |
+
import time
|
| 21 |
+
|
| 22 |
+
PROTOCOL = "looprl-math-pass8-16k-v1"
|
| 23 |
+
DATASETS = ("aime24", "aime25", "aime26", "math500", "amc23", "olympiadbench",
|
| 24 |
+
"minerva_math", "hmmt_feb25", "hmmt_nov25")
|
| 25 |
+
K, MAX_TOKENS, CONCURRENCY, BATCHED_TOKENS = 8, 16384, 8, 4096
|
| 26 |
+
TEMPERATURE, TOP_P, MEMORY = 0.6, 0.95, 0.60
|
| 27 |
+
INSTRUCTION = "Please reason step by step, and put your final answer within \\boxed{}."
|
| 28 |
+
GRADER = "last-complete-box-plain-or-math-verify-v1"
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def digest(path):
|
| 32 |
+
h = hashlib.sha256()
|
| 33 |
+
with Path(path).open("rb") as f:
|
| 34 |
+
for block in iter(lambda: f.read(4 << 20), b""):
|
| 35 |
+
h.update(block)
|
| 36 |
+
return h.hexdigest()
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def write_json(path, value):
|
| 40 |
+
path = Path(path)
|
| 41 |
+
path.parent.mkdir(parents=True, exist_ok=True)
|
| 42 |
+
tmp = path.with_name(path.name + f".{os.getpid()}.tmp")
|
| 43 |
+
tmp.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n")
|
| 44 |
+
os.replace(tmp, path)
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def last_boxed(text):
|
| 48 |
+
"""Last *complete* boxed answer; never fall back to an arbitrary number."""
|
| 49 |
+
answer = None
|
| 50 |
+
for match in re.finditer(r"\\(?:boxed|fbox)\s*\{", text or ""):
|
| 51 |
+
depth, pos = 1, match.end()
|
| 52 |
+
while pos < len(text) and depth:
|
| 53 |
+
depth += (text[pos] == "{") - (text[pos] == "}")
|
| 54 |
+
pos += 1
|
| 55 |
+
if depth == 0:
|
| 56 |
+
answer = text[match.end():pos - 1].strip()
|
| 57 |
+
return answer
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def clean_math(value):
|
| 61 |
+
text = str(value).strip()
|
| 62 |
+
if text.startswith("$") and text.endswith("$"):
|
| 63 |
+
text = text.strip("$").strip()
|
| 64 |
+
return text
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def grade(gold, predicted):
|
| 68 |
+
if predicted is None or not str(predicted).strip():
|
| 69 |
+
return False
|
| 70 |
+
gold, predicted = clean_math(gold), clean_math(predicted)
|
| 71 |
+
# Preserve case, percent signs, interval endpoints and mathematical operators.
|
| 72 |
+
if re.sub(r"\s+", "", gold) == re.sub(r"\s+", "", predicted):
|
| 73 |
+
return True
|
| 74 |
+
from math_verify import parse, verify
|
| 75 |
+
try:
|
| 76 |
+
g = parse(f"${gold}$", parsing_timeout=1)
|
| 77 |
+
p = parse(f"${predicted}$", parsing_timeout=1)
|
| 78 |
+
return bool(g) and bool(p) and bool(verify(g, p, timeout_seconds=1))
|
| 79 |
+
except Exception:
|
| 80 |
+
return False
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
def sample_seed(uid, replicate):
|
| 84 |
+
value = f"{PROTOCOL}|{uid}|{replicate}|42".encode()
|
| 85 |
+
return int(hashlib.sha256(value).hexdigest()[:16], 16) & 0x7FFFFFFF
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
def normalize_row(dataset, row, index):
|
| 89 |
+
question = row.get("question") or row.get("problem")
|
| 90 |
+
if row.get("context"):
|
| 91 |
+
question = str(row["context"]) + "\n\n" + question
|
| 92 |
+
gold = row.get("answer")
|
| 93 |
+
if gold is None:
|
| 94 |
+
gold = row.get("final_answer")
|
| 95 |
+
if isinstance(gold, list):
|
| 96 |
+
# Multiple components are required jointly, not alternative correct answers.
|
| 97 |
+
gold = ", ".join(clean_math(x) for x in gold)
|
| 98 |
+
if gold is None:
|
| 99 |
+
solution = row.get("solution", "")
|
| 100 |
+
if isinstance(solution, list):
|
| 101 |
+
solution = "\n".join(solution)
|
| 102 |
+
gold = last_boxed(solution)
|
| 103 |
+
if isinstance(gold, float) and math.isfinite(gold):
|
| 104 |
+
gold = format(gold, ".15g")
|
| 105 |
+
if not question or gold is None or not str(gold).strip():
|
| 106 |
+
raise ValueError(f"Missing question/gold in {dataset} row {index}")
|
| 107 |
+
source_id = next((row[k] for k in ("id", "idx", "problem_idx")
|
| 108 |
+
if row.get(k) is not None), index)
|
| 109 |
+
return {"dataset": dataset, "uid": f"{dataset}/{source_id}",
|
| 110 |
+
"source_index": index, "question": question, "gold": clean_math(gold)}
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def load_rows(root):
|
| 114 |
+
path = root / "data/questions.jsonl"
|
| 115 |
+
manifest = json.loads((root / "protocol.json").read_text())
|
| 116 |
+
if digest(path) != manifest["questions_sha256"]:
|
| 117 |
+
raise ValueError("Frozen evaluation questions changed")
|
| 118 |
+
return [json.loads(l) for l in path.read_text().splitlines() if l.strip()]
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
def execution_parameters(protocol):
|
| 122 |
+
concurrency = protocol["concurrency_per_gpu"]
|
| 123 |
+
tokens = protocol["max_num_batched_tokens"]
|
| 124 |
+
memory = protocol["gpu_memory_utilization"]
|
| 125 |
+
if type(concurrency) is not int or concurrency < 1:
|
| 126 |
+
raise ValueError("Invalid concurrency_per_gpu")
|
| 127 |
+
if type(tokens) is not int or tokens < concurrency:
|
| 128 |
+
raise ValueError("Invalid max_num_batched_tokens")
|
| 129 |
+
if not isinstance(memory, (int, float)) or not 0 < memory < 1:
|
| 130 |
+
raise ValueError("Invalid gpu_memory_utilization")
|
| 131 |
+
return concurrency, tokens, float(memory)
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
def journal_records(path, *, allow_incomplete=False):
|
| 135 |
+
"""Read complete records; a killed writer may leave only its final line partial."""
|
| 136 |
+
with Path(path).open("rb") as f:
|
| 137 |
+
for line in f:
|
| 138 |
+
if not line.endswith(b"\n"):
|
| 139 |
+
if allow_incomplete:
|
| 140 |
+
return
|
| 141 |
+
raise ValueError(f"Incomplete journal tail: {path}")
|
| 142 |
+
yield json.loads(line)
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def validate_trial(record, known):
|
| 146 |
+
uid, rep = record["uid"], record["replicate"]
|
| 147 |
+
if uid not in known or type(rep) is not int or not 0 <= rep < K:
|
| 148 |
+
raise ValueError("Unknown committed trial")
|
| 149 |
+
row = known[uid]
|
| 150 |
+
if (record["protocol"] != PROTOCOL or record["seed"] != sample_seed(uid, rep)
|
| 151 |
+
or record["dataset"] != row["dataset"] or record["gold"] != row["gold"]
|
| 152 |
+
or record["prompt_tokens"] != len(row["prompt_token_ids"])
|
| 153 |
+
or type(record["correct"]) is not bool
|
| 154 |
+
or record["generated_tokens"] != len(record["token_ids"])
|
| 155 |
+
or not 0 < record["generated_tokens"] <= MAX_TOKENS):
|
| 156 |
+
raise ValueError("Committed trial does not match the frozen evaluation")
|
| 157 |
+
return uid, rep
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
def prepare(args):
|
| 161 |
+
import pyarrow.parquet as pq
|
| 162 |
+
from transformers import AutoTokenizer
|
| 163 |
+
root, model = args.root, args.model.resolve()
|
| 164 |
+
if (root / "protocol.json").exists():
|
| 165 |
+
raise FileExistsError("Prepared protocol exists; do not overwrite it")
|
| 166 |
+
sources = json.loads((root / "data/sources.json").read_text())
|
| 167 |
+
assert set(sources) == set(DATASETS)
|
| 168 |
+
by_dataset = {}
|
| 169 |
+
for name in DATASETS:
|
| 170 |
+
src = sources[name]
|
| 171 |
+
path = Path(src["raw_file"])
|
| 172 |
+
assert digest(path) == src["raw_sha256"], name
|
| 173 |
+
raw = (pq.read_table(path).to_pylist() if path.suffix == ".parquet" else
|
| 174 |
+
[json.loads(l) for l in path.read_text().splitlines() if l.strip()])
|
| 175 |
+
assert len(raw) == src["rows"], name
|
| 176 |
+
by_dataset[name] = [normalize_row(name, row, i) for i, row in enumerate(raw)]
|
| 177 |
+
assert len({r["uid"] for r in by_dataset[name]}) == len(raw), name
|
| 178 |
+
# Interleave the suites so small datasets are not held behind all of MATH-500.
|
| 179 |
+
rows = [by_dataset[name][i] for i in range(max(map(len, by_dataset.values())))
|
| 180 |
+
for name in DATASETS if i < len(by_dataset[name])]
|
| 181 |
+
tok = AutoTokenizer.from_pretrained(str(model), trust_remote_code=True,
|
| 182 |
+
local_files_only=True)
|
| 183 |
+
for r in rows:
|
| 184 |
+
prompt = tok.apply_chat_template(
|
| 185 |
+
[{"role": "user", "content": r["question"] + "\n\n" + INSTRUCTION}],
|
| 186 |
+
tokenize=False, add_generation_prompt=True, enable_thinking=True)
|
| 187 |
+
r["prompt_token_ids"] = tok.encode(prompt, add_special_tokens=False)
|
| 188 |
+
r["prompt_sha256"] = hashlib.sha256(prompt.encode()).hexdigest()
|
| 189 |
+
max_prompt = max(len(r["prompt_token_ids"]) for r in rows)
|
| 190 |
+
context = math.ceil((max_prompt + MAX_TOKENS) / 256) * 256
|
| 191 |
+
if context > 32768:
|
| 192 |
+
raise ValueError("Prompt plus 16k exceeds the model's 32k SFT context")
|
| 193 |
+
data = root / "data/questions.jsonl"
|
| 194 |
+
data.write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows))
|
| 195 |
+
cfg = json.loads((model / "config.json").read_text())
|
| 196 |
+
assert cfg["total_ut_steps"] == 4 and cfg["num_hidden_layers"] == 24
|
| 197 |
+
provenance = json.loads((model / "provenance.json").read_text())
|
| 198 |
+
assert provenance["model_id"] == "ByteDance/Ouro-1.4B-Thinking"
|
| 199 |
+
print("CHECKING_BASE_WEIGHTS", flush=True)
|
| 200 |
+
weights_sha = digest(model / "model.safetensors")
|
| 201 |
+
assert weights_sha == provenance["files_sha256"]["model.safetensors"]
|
| 202 |
+
protocol = {"protocol": PROTOCOL, "model": str(model),
|
| 203 |
+
"model_id": provenance["model_id"], "revision": provenance["revision"],
|
| 204 |
+
"weights_sha256": weights_sha, "questions_sha256": digest(data),
|
| 205 |
+
"datasets": {k: len(v) for k, v in by_dataset.items()},
|
| 206 |
+
"questions": len(rows), "responses": K * len(rows),
|
| 207 |
+
"k": K, "max_new_tokens": MAX_TOKENS, "max_prompt_tokens": max_prompt,
|
| 208 |
+
"max_model_len": context, "loops": 4, "dtype": "bfloat16",
|
| 209 |
+
"temperature": TEMPERATURE, "top_p": TOP_P, "top_k": -1,
|
| 210 |
+
"concurrency_per_gpu": CONCURRENCY, "max_num_batched_tokens": BATCHED_TOKENS,
|
| 211 |
+
"gpu_memory_utilization": MEMORY, "tensor_parallel_size": 1,
|
| 212 |
+
"prompt": "zero-shot, Thinking chat template, training math instruction",
|
| 213 |
+
"grader": GRADER, "length_stops": "recorded; scored by the same last-box rule",
|
| 214 |
+
"source_sha256": digest(Path(__file__)),
|
| 215 |
+
"versions": {x: importlib.metadata.version(x) for x in
|
| 216 |
+
("vllm", "torch", "transformers", "math-verify")}}
|
| 217 |
+
write_json(root / "protocol.json", protocol)
|
| 218 |
+
print("PREPARED", json.dumps(protocol), flush=True)
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
def summary(rows, records):
|
| 222 |
+
known = {r["uid"]: r for r in rows}
|
| 223 |
+
grouped = collections.defaultdict(dict)
|
| 224 |
+
for record in records:
|
| 225 |
+
uid, rep = record["uid"], record["replicate"]
|
| 226 |
+
if uid not in known or not isinstance(rep, int) or not 0 <= rep < K:
|
| 227 |
+
raise ValueError("Unknown evaluation trial")
|
| 228 |
+
if rep in grouped[uid]:
|
| 229 |
+
raise ValueError("Duplicate evaluation trial")
|
| 230 |
+
grouped[uid][rep] = record
|
| 231 |
+
suites = {}
|
| 232 |
+
for name in DATASETS:
|
| 233 |
+
questions = [r for r in rows if r["dataset"] == name]
|
| 234 |
+
complete = [grouped[r["uid"]] for r in questions if len(grouped[r["uid"]]) == K]
|
| 235 |
+
correct = [sum(bool(v["correct"]) for v in g.values()) for g in complete]
|
| 236 |
+
suites[name] = {"questions": len(questions), "completed_questions": len(complete),
|
| 237 |
+
"pass_at_8": sum(c > 0 for c in correct) / len(complete) if complete else None,
|
| 238 |
+
"mean_sample_accuracy": sum(correct) / (K * len(complete)) if complete else None,
|
| 239 |
+
"status": "complete" if len(complete) == len(questions) else "partial"}
|
| 240 |
+
return {"protocol": PROTOCOL, "responses_completed": len(records),
|
| 241 |
+
"responses_expected": len(rows) * K,
|
| 242 |
+
"status": "complete" if len(records) == len(rows) * K else "partial",
|
| 243 |
+
"datasets": suites}
|
| 244 |
+
|
| 245 |
+
|
| 246 |
+
def summarize(root):
|
| 247 |
+
rows = load_rows(root)
|
| 248 |
+
records = []
|
| 249 |
+
for path in sorted((root / "responses").glob("rank*.jsonl")):
|
| 250 |
+
records.extend(journal_records(path, allow_incomplete=True))
|
| 251 |
+
result = summary(rows, records)
|
| 252 |
+
write_json(root / "summary.json", result)
|
| 253 |
+
print("SUMMARY", json.dumps(result), flush=True)
|
| 254 |
+
return result
|
| 255 |
+
|
| 256 |
+
|
| 257 |
+
def worker(args):
|
| 258 |
+
from vllm import LLM, SamplingParams
|
| 259 |
+
from vllm.engine.arg_utils import EngineArgs
|
| 260 |
+
from vllm.sampling_params import RequestOutputKind
|
| 261 |
+
root = args.root
|
| 262 |
+
lock = root / "locks" / f"rank{args.rank}.lock"
|
| 263 |
+
lock.parent.mkdir(exist_ok=True)
|
| 264 |
+
lockfile = lock.open("a")
|
| 265 |
+
fcntl.flock(lockfile, fcntl.LOCK_EX | fcntl.LOCK_NB)
|
| 266 |
+
protocol = json.loads((root / "protocol.json").read_text())
|
| 267 |
+
assert protocol["source_sha256"] == digest(Path(__file__))
|
| 268 |
+
concurrency, batched_tokens, memory = execution_parameters(protocol)
|
| 269 |
+
rows = load_rows(root)[args.rank::args.world_size]
|
| 270 |
+
known = {r["uid"]: r for r in rows}
|
| 271 |
+
journal = root / "responses" / f"rank{args.rank}.jsonl"
|
| 272 |
+
journal.parent.mkdir(exist_ok=True)
|
| 273 |
+
done = set()
|
| 274 |
+
if journal.exists():
|
| 275 |
+
for record in journal_records(journal):
|
| 276 |
+
key = validate_trial(record, known)
|
| 277 |
+
if key in done:
|
| 278 |
+
raise ValueError("Duplicate committed trial")
|
| 279 |
+
done.add(key)
|
| 280 |
+
resumed = len(done)
|
| 281 |
+
todo = collections.deque((r, rep) for r in rows for rep in range(K)
|
| 282 |
+
if (r["uid"], rep) not in done)
|
| 283 |
+
kwargs = dict(model=protocol["model"], tokenizer=protocol["model"],
|
| 284 |
+
trust_remote_code=True, dtype="bfloat16", tensor_parallel_size=1,
|
| 285 |
+
distributed_executor_backend="mp", hf_overrides={"total_ut_steps": 4},
|
| 286 |
+
gpu_memory_utilization=memory, max_model_len=protocol["max_model_len"],
|
| 287 |
+
max_num_seqs=concurrency, max_num_batched_tokens=batched_tokens,
|
| 288 |
+
enable_chunked_prefill=True, enable_prefix_caching=False,
|
| 289 |
+
generation_config="vllm", seed=42, disable_log_stats=False)
|
| 290 |
+
fields = {f.name for f in dataclasses.fields(EngineArgs)}
|
| 291 |
+
for key, value in {"model_impl": "vllm", "async_scheduling": False}.items():
|
| 292 |
+
if key in fields:
|
| 293 |
+
kwargs[key] = value
|
| 294 |
+
llm = LLM(**kwargs)
|
| 295 |
+
engine, tokenizer = llm.llm_engine, llm.get_tokenizer()
|
| 296 |
+
config = engine.vllm_config
|
| 297 |
+
assert config.model_config.hf_config.total_ut_steps == 4
|
| 298 |
+
assert config.scheduler_config.max_num_seqs == concurrency
|
| 299 |
+
assert config.scheduler_config.max_num_batched_tokens == batched_tokens
|
| 300 |
+
cache = config.cache_config
|
| 301 |
+
capacity = cache.num_gpu_blocks * cache.block_size
|
| 302 |
+
if capacity < concurrency * protocol["max_model_len"]:
|
| 303 |
+
raise RuntimeError(f"KV cache cannot hold all {concurrency} full-budget sequences")
|
| 304 |
+
identity = {"rank": args.rank, "world_size": args.world_size, "pid": os.getpid(),
|
| 305 |
+
"cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES"),
|
| 306 |
+
"engine_kwargs": kwargs, "kv_capacity_tokens": capacity,
|
| 307 |
+
"worst_case_active_tokens": concurrency * protocol["max_model_len"],
|
| 308 |
+
"responses_resumed": resumed, "execution_id": root.name}
|
| 309 |
+
write_json(root / "preflight" / f"engine_rank{args.rank}.json", identity)
|
| 310 |
+
print("ENGINE_READY", json.dumps(identity), flush=True)
|
| 311 |
+
active = {}
|
| 312 |
+
started, last_progress = time.monotonic(), 0.0
|
| 313 |
+
generated_tokens = 0
|
| 314 |
+
progress_file = root / "workers" / f"rank{args.rank}.json"
|
| 315 |
+
with journal.open("a", buffering=1) as output:
|
| 316 |
+
while todo or active:
|
| 317 |
+
while todo and len(active) < concurrency:
|
| 318 |
+
row, rep = todo.popleft()
|
| 319 |
+
key = f"{row['uid']}:{rep}"
|
| 320 |
+
params = SamplingParams(n=1, temperature=TEMPERATURE, top_p=TOP_P,
|
| 321 |
+
top_k=-1, max_tokens=MAX_TOKENS,
|
| 322 |
+
seed=sample_seed(row["uid"], rep),
|
| 323 |
+
output_kind=RequestOutputKind.FINAL_ONLY)
|
| 324 |
+
engine.add_request(key, {"prompt_token_ids": row["prompt_token_ids"]}, params)
|
| 325 |
+
active[key] = (row, rep)
|
| 326 |
+
for request in engine.step():
|
| 327 |
+
if not request.finished:
|
| 328 |
+
continue
|
| 329 |
+
row, rep = active.pop(request.request_id)
|
| 330 |
+
assert len(request.outputs) == 1
|
| 331 |
+
response = request.outputs[0]
|
| 332 |
+
text = response.text or tokenizer.decode(response.token_ids,
|
| 333 |
+
skip_special_tokens=False)
|
| 334 |
+
prediction = last_boxed(text)
|
| 335 |
+
record = {"protocol": PROTOCOL, "uid": row["uid"],
|
| 336 |
+
"dataset": row["dataset"], "replicate": rep,
|
| 337 |
+
"seed": sample_seed(row["uid"], rep), "gold": row["gold"],
|
| 338 |
+
"prediction": prediction, "correct": grade(row["gold"], prediction),
|
| 339 |
+
"response": text, "token_ids": list(response.token_ids),
|
| 340 |
+
"generated_tokens": len(response.token_ids),
|
| 341 |
+
"prompt_tokens": len(row["prompt_token_ids"]),
|
| 342 |
+
"finish_reason": response.finish_reason,
|
| 343 |
+
"stop_reason": response.stop_reason,
|
| 344 |
+
"finished_unix": time.time(), "rank": args.rank,
|
| 345 |
+
"execution_id": root.name}
|
| 346 |
+
assert 0 < record["generated_tokens"] <= MAX_TOKENS
|
| 347 |
+
output.write(json.dumps(record, ensure_ascii=False) + "\n")
|
| 348 |
+
output.flush()
|
| 349 |
+
os.fsync(output.fileno())
|
| 350 |
+
done.add((row["uid"], rep))
|
| 351 |
+
generated_tokens += record["generated_tokens"]
|
| 352 |
+
print(f"RESPONSE uid={row['uid']} rep={rep} tokens={record['generated_tokens']} "
|
| 353 |
+
f"correct={record['correct']} reason={response.finish_reason}", flush=True)
|
| 354 |
+
now = time.monotonic()
|
| 355 |
+
if now - last_progress >= 20 or not (todo or active):
|
| 356 |
+
progress = {**identity, "status": "running" if (todo or active) else "complete",
|
| 357 |
+
"active_sequences": len(active), "queued_sequences": len(todo),
|
| 358 |
+
"responses_completed": len(done), "responses_expected": len(rows) * K,
|
| 359 |
+
"generated_tokens_this_session": generated_tokens,
|
| 360 |
+
"elapsed_seconds": now - started, "updated_unix": time.time()}
|
| 361 |
+
write_json(progress_file, progress)
|
| 362 |
+
print("PROGRESS", json.dumps(progress), flush=True)
|
| 363 |
+
last_progress = now
|
| 364 |
+
write_json(root / "workers" / f"rank{args.rank}.complete.json",
|
| 365 |
+
{"responses": len(done), "expected": len(rows) * K, "finished_unix": time.time()})
|
| 366 |
+
engine.engine_core.shutdown()
|
| 367 |
+
|
| 368 |
+
|
| 369 |
+
def main():
|
| 370 |
+
parser = argparse.ArgumentParser(description=__doc__)
|
| 371 |
+
parser.add_argument("mode", choices=("prepare", "worker", "summarize"))
|
| 372 |
+
parser.add_argument("--root", type=Path, required=True)
|
| 373 |
+
parser.add_argument("--model", type=Path)
|
| 374 |
+
parser.add_argument("--rank", type=int, default=0)
|
| 375 |
+
parser.add_argument("--world-size", type=int, default=7)
|
| 376 |
+
args = parser.parse_args()
|
| 377 |
+
if args.mode == "prepare":
|
| 378 |
+
if args.model is None:
|
| 379 |
+
parser.error("prepare requires --model")
|
| 380 |
+
prepare(args)
|
| 381 |
+
elif args.mode == "worker":
|
| 382 |
+
if not 0 <= args.rank < args.world_size:
|
| 383 |
+
parser.error("rank is outside world size")
|
| 384 |
+
worker(args)
|
| 385 |
+
else:
|
| 386 |
+
summarize(args.root)
|
| 387 |
+
|
| 388 |
+
|
| 389 |
+
if __name__ == "__main__":
|
| 390 |
+
main()
|
evaluations/base_pass8_16k_seed42/code/scripts/resume_math_eval.py
ADDED
|
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Copy committed trials from a stopped evaluation into one new execution."""
|
| 3 |
+
import argparse
|
| 4 |
+
from contextlib import ExitStack
|
| 5 |
+
import fcntl
|
| 6 |
+
import hashlib
|
| 7 |
+
import json
|
| 8 |
+
import os
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
import shutil
|
| 11 |
+
import sys
|
| 12 |
+
import time
|
| 13 |
+
|
| 14 |
+
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
| 15 |
+
from rltt_baseline.eval_math_suite import (
|
| 16 |
+
PROTOCOL, digest, execution_parameters, load_rows, validate_trial, write_json,
|
| 17 |
+
)
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def copy_committed(source, target, known, rank):
|
| 21 |
+
"""Preserve exact bytes, rejecting duplicates and any non-final damaged record."""
|
| 22 |
+
seen, total, incomplete, h = set(), 0, 0, hashlib.sha256()
|
| 23 |
+
if source.stat().st_size > 2 << 30:
|
| 24 |
+
raise ValueError("Source journal exceeds bounded migration size")
|
| 25 |
+
with source.open("rb") as inp, target.open("xb") as out:
|
| 26 |
+
for line in inp:
|
| 27 |
+
if not line.endswith(b"\n"):
|
| 28 |
+
incomplete = len(line)
|
| 29 |
+
break
|
| 30 |
+
record = json.loads(line)
|
| 31 |
+
key = validate_trial(record, known)
|
| 32 |
+
if record["rank"] != rank or key in seen:
|
| 33 |
+
raise ValueError("Duplicate or incorrectly sharded committed trial")
|
| 34 |
+
seen.add(key)
|
| 35 |
+
out.write(line)
|
| 36 |
+
h.update(line)
|
| 37 |
+
total += len(line)
|
| 38 |
+
out.flush()
|
| 39 |
+
os.fsync(out.fileno())
|
| 40 |
+
return {"rank": rank, "responses": len(seen), "bytes": total,
|
| 41 |
+
"committed_sha256": h.hexdigest(), "incomplete_tail_bytes": incomplete}
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def resume(source, target, concurrency, batched_tokens, memory):
|
| 45 |
+
source, target = source.resolve(), target.resolve()
|
| 46 |
+
if source == target or (target / "protocol.json").exists():
|
| 47 |
+
raise ValueError("Refuse to overwrite a frozen execution")
|
| 48 |
+
src_protocol = json.loads((source / "protocol.json").read_text())
|
| 49 |
+
if src_protocol["protocol"] != PROTOCOL:
|
| 50 |
+
raise ValueError("Unknown source evaluation protocol")
|
| 51 |
+
if digest(source / "code/rltt_baseline/eval_math_suite.py") != src_protocol["source_sha256"]:
|
| 52 |
+
raise ValueError("Source code snapshot changed")
|
| 53 |
+
launch = json.loads((source / "preflight/launch_receipt.json").read_text())
|
| 54 |
+
world_size = launch["workers"]
|
| 55 |
+
if world_size != 7 or launch["gpus"] != list(range(1, 8)):
|
| 56 |
+
raise ValueError("Refuse to change existing evaluation sharding")
|
| 57 |
+
protocol = dict(src_protocol)
|
| 58 |
+
protocol.update(concurrency_per_gpu=concurrency, max_num_batched_tokens=batched_tokens,
|
| 59 |
+
gpu_memory_utilization=memory)
|
| 60 |
+
execution_parameters(protocol)
|
| 61 |
+
rows = load_rows(source)
|
| 62 |
+
# Locks are held for the complete copy. No live coordinator or worker may write.
|
| 63 |
+
with ExitStack() as stack:
|
| 64 |
+
for path in [source / "coordinator.lock"] + [
|
| 65 |
+
source / "locks" / f"rank{i}.lock" for i in range(world_size)]:
|
| 66 |
+
handle = stack.enter_context(path.open("r"))
|
| 67 |
+
fcntl.flock(handle, fcntl.LOCK_EX | fcntl.LOCK_NB)
|
| 68 |
+
for name in ["data", "responses", "preflight", "workers", "logs", "locks"]:
|
| 69 |
+
(target / name).mkdir(parents=True, exist_ok=True)
|
| 70 |
+
shutil.copyfile(source / "data/questions.jsonl", target / "data/questions.jsonl")
|
| 71 |
+
shutil.copyfile(source / "data/sources.json", target / "data/sources.json")
|
| 72 |
+
journals = []
|
| 73 |
+
for rank in range(world_size):
|
| 74 |
+
known = {r["uid"]: r for r in rows[rank::world_size]}
|
| 75 |
+
journals.append(copy_committed(source / "responses" / f"rank{rank}.jsonl",
|
| 76 |
+
target / "responses" / f"rank{rank}.jsonl", known, rank))
|
| 77 |
+
if digest(target / "data/questions.jsonl") != src_protocol["questions_sha256"]:
|
| 78 |
+
raise ValueError("Question copy failed identity check")
|
| 79 |
+
protocol.update(source_sha256=digest(target / "code/rltt_baseline/eval_math_suite.py"),
|
| 80 |
+
parent_root=str(source), execution_id=target.name,
|
| 81 |
+
responses_resumed=sum(j["responses"] for j in journals))
|
| 82 |
+
info = {"parent_root": str(source), "new_root": str(target),
|
| 83 |
+
"source_protocol_sha256": digest(source / "protocol.json"),
|
| 84 |
+
"questions_sha256": protocol["questions_sha256"],
|
| 85 |
+
"model_weights_sha256": protocol["weights_sha256"],
|
| 86 |
+
"resumed_responses": protocol["responses_resumed"],
|
| 87 |
+
"world_size": world_size, "gpus": launch["gpus"],
|
| 88 |
+
"runtime_before": {k: src_protocol[k] for k in ["concurrency_per_gpu",
|
| 89 |
+
"max_num_batched_tokens", "gpu_memory_utilization"]},
|
| 90 |
+
"runtime_after": {k: protocol[k] for k in ["concurrency_per_gpu",
|
| 91 |
+
"max_num_batched_tokens", "gpu_memory_utilization"]},
|
| 92 |
+
"sampling_identity_unchanged": True, "journals": journals,
|
| 93 |
+
"prepared_unix": time.time(), "status": "prepared_not_started"}
|
| 94 |
+
write_json(target / "preflight/resume.json", info)
|
| 95 |
+
write_json(target / "protocol.json", protocol)
|
| 96 |
+
return info
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
def main():
|
| 100 |
+
p = argparse.ArgumentParser(description=__doc__)
|
| 101 |
+
p.add_argument("--from-root", type=Path, required=True)
|
| 102 |
+
p.add_argument("--root", type=Path, required=True)
|
| 103 |
+
p.add_argument("--concurrency", type=int, required=True)
|
| 104 |
+
p.add_argument("--max-num-batched-tokens", type=int, required=True)
|
| 105 |
+
p.add_argument("--gpu-memory-utilization", type=float, required=True)
|
| 106 |
+
a = p.parse_args()
|
| 107 |
+
print(json.dumps(resume(a.from_root, a.root, a.concurrency,
|
| 108 |
+
a.max_num_batched_tokens, a.gpu_memory_utilization), indent=2), flush=True)
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
if __name__ == "__main__":
|
| 112 |
+
main()
|
evaluations/base_pass8_16k_seed42/code/scripts/run_math_eval_suite.py
ADDED
|
@@ -0,0 +1,131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Launch one fixed-policy math suite on explicitly listed idle GPUs, once."""
|
| 3 |
+
import argparse
|
| 4 |
+
import fcntl
|
| 5 |
+
import hashlib
|
| 6 |
+
import json
|
| 7 |
+
import os
|
| 8 |
+
from pathlib import Path
|
| 9 |
+
import subprocess
|
| 10 |
+
import sys
|
| 11 |
+
import time
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def save(path, value):
|
| 15 |
+
tmp = path.with_suffix(".tmp")
|
| 16 |
+
tmp.write_text(json.dumps(value, indent=2) + "\n")
|
| 17 |
+
os.replace(tmp, path)
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def main():
|
| 21 |
+
p = argparse.ArgumentParser(description=__doc__)
|
| 22 |
+
p.add_argument("--root", type=Path, required=True)
|
| 23 |
+
p.add_argument("--gpus", required=True, help="Physical GPU indices, comma-separated")
|
| 24 |
+
args = p.parse_args()
|
| 25 |
+
root = args.root.resolve()
|
| 26 |
+
gpus = [int(v) for v in args.gpus.split(",")]
|
| 27 |
+
if len(set(gpus)) != len(gpus) or not gpus:
|
| 28 |
+
p.error("GPU indices must be distinct")
|
| 29 |
+
lockfile = (root / "coordinator.lock").open("a")
|
| 30 |
+
fcntl.flock(lockfile, fcntl.LOCK_EX | fcntl.LOCK_NB)
|
| 31 |
+
receipt = root / "preflight/launch_receipt.json"
|
| 32 |
+
if receipt.exists():
|
| 33 |
+
raise RuntimeError("A launch receipt exists; inspect this same attempt before relaunching")
|
| 34 |
+
protocol = json.loads((root / "protocol.json").read_text())
|
| 35 |
+
script = root / "code/rltt_baseline/eval_math_suite.py"
|
| 36 |
+
if hashlib.sha256(script.read_bytes()).hexdigest() != protocol["source_sha256"]:
|
| 37 |
+
raise RuntimeError("Frozen evaluation code changed")
|
| 38 |
+
inventory = subprocess.check_output(
|
| 39 |
+
["nvidia-smi", "--query-gpu=index,uuid", "--format=csv,noheader"], text=True)
|
| 40 |
+
uuids = {int(l.split(",")[0]): l.split(",")[1].strip() for l in inventory.splitlines()}
|
| 41 |
+
apps = subprocess.check_output(
|
| 42 |
+
["nvidia-smi", "--query-compute-apps=gpu_uuid,pid", "--format=csv,noheader"], text=True)
|
| 43 |
+
occupied = {l.split(",")[0].strip() for l in apps.splitlines() if "," in l}
|
| 44 |
+
for gpu in gpus:
|
| 45 |
+
if gpu not in uuids or uuids[gpu] in occupied:
|
| 46 |
+
raise RuntimeError(f"Requested GPU {gpu} is unavailable or already has a compute process")
|
| 47 |
+
(root / "logs").mkdir(exist_ok=True)
|
| 48 |
+
info = {"logical_id": root.name, "pid": os.getpid(), "gpus": gpus,
|
| 49 |
+
"workers": len(gpus), "concurrency_per_gpu": protocol["concurrency_per_gpu"],
|
| 50 |
+
"global_concurrency_limit": len(gpus) * protocol["concurrency_per_gpu"],
|
| 51 |
+
"protocol_sha256": hashlib.sha256((root / "protocol.json").read_bytes()).hexdigest(),
|
| 52 |
+
"started_unix": time.time(), "status": "starting", "children": []}
|
| 53 |
+
save(receipt, info)
|
| 54 |
+
children, logs = [], []
|
| 55 |
+
|
| 56 |
+
def start(rank):
|
| 57 |
+
gpu = gpus[rank]
|
| 58 |
+
env = os.environ.copy()
|
| 59 |
+
env.update(CUDA_VISIBLE_DEVICES=str(gpu), OMP_NUM_THREADS="4",
|
| 60 |
+
TOKENIZERS_PARALLELISM="false", PYTHONDONTWRITEBYTECODE="1",
|
| 61 |
+
HF_HUB_OFFLINE="1", TRANSFORMERS_OFFLINE="1", HF_DATASETS_OFFLINE="1",
|
| 62 |
+
TRITON_PTXAS_PATH="/usr/local/cuda/bin/ptxas",
|
| 63 |
+
VLLM_WORKER_MULTIPROC_METHOD="spawn")
|
| 64 |
+
cache = root / "cache" / f"rank{rank}"
|
| 65 |
+
env["VLLM_CACHE_ROOT"] = str(cache / "vllm")
|
| 66 |
+
env["TORCHINDUCTOR_CACHE_DIR"] = str(cache / "inductor")
|
| 67 |
+
log = (root / "logs" / f"rank{rank}.log").open("a", buffering=1)
|
| 68 |
+
logs.append(log)
|
| 69 |
+
cmd = [sys.executable, "-u", str(script), "worker", "--root", str(root),
|
| 70 |
+
"--rank", str(rank), "--world-size", str(len(gpus))]
|
| 71 |
+
child = subprocess.Popen(cmd, env=env, stdout=log, stderr=subprocess.STDOUT,
|
| 72 |
+
cwd="/opt/looprl-work/LoopRL")
|
| 73 |
+
children.append(child)
|
| 74 |
+
info["children"].append({"rank": rank, "gpu": gpu, "pid": child.pid,
|
| 75 |
+
"log": str(root / "logs" / f"rank{rank}.log")})
|
| 76 |
+
save(receipt, info)
|
| 77 |
+
print(f"WORKER_SUBMITTED rank={rank} gpu={gpu} pid={child.pid}", flush=True)
|
| 78 |
+
|
| 79 |
+
start(0)
|
| 80 |
+
# Validate the cache budget and a newly committed response before using all GPUs.
|
| 81 |
+
deadline = time.monotonic() + 900
|
| 82 |
+
while True:
|
| 83 |
+
if children[0].poll() is not None:
|
| 84 |
+
if children[0].returncode != 0:
|
| 85 |
+
info["status"] = "initial_worker_failed"
|
| 86 |
+
info["exit_codes"] = [children[0].returncode]
|
| 87 |
+
save(receipt, info)
|
| 88 |
+
(root / "eval_exit_code").write_text("1\n")
|
| 89 |
+
return 1
|
| 90 |
+
engine_file = root / "preflight/engine_rank0.json"
|
| 91 |
+
progress_file = root / "workers/rank0.json"
|
| 92 |
+
if engine_file.exists() and progress_file.exists():
|
| 93 |
+
engine = json.loads(engine_file.read_text())
|
| 94 |
+
progress = json.loads(progress_file.read_text())
|
| 95 |
+
if (progress["responses_completed"] > engine.get("responses_resumed", 0)
|
| 96 |
+
or progress["status"] == "complete"):
|
| 97 |
+
info["first_worker_verified"] = {
|
| 98 |
+
"responses_resumed": engine.get("responses_resumed", 0),
|
| 99 |
+
"responses_completed": progress["responses_completed"],
|
| 100 |
+
"kv_capacity_tokens": engine["kv_capacity_tokens"],
|
| 101 |
+
}
|
| 102 |
+
save(receipt, info)
|
| 103 |
+
break
|
| 104 |
+
if time.monotonic() > deadline:
|
| 105 |
+
info["status"] = "initial_worker_startup_unconfirmed"
|
| 106 |
+
save(receipt, info)
|
| 107 |
+
return 2 # Preserve the one worker; do not duplicate or kill an unknown launch.
|
| 108 |
+
time.sleep(1)
|
| 109 |
+
for rank in range(1, len(gpus)):
|
| 110 |
+
start(rank)
|
| 111 |
+
info["status"] = "running"
|
| 112 |
+
save(receipt, info)
|
| 113 |
+
while any(c.poll() is None for c in children):
|
| 114 |
+
time.sleep(5)
|
| 115 |
+
codes = [c.wait() for c in children]
|
| 116 |
+
for rank, code in enumerate(codes):
|
| 117 |
+
(root / "workers" / f"rank{rank}.exit_code").write_text(f"{code}\n")
|
| 118 |
+
cmd = [sys.executable, str(script), "summarize", "--root", str(root)]
|
| 119 |
+
summed = subprocess.run(cmd, check=False)
|
| 120 |
+
result = json.loads((root / "summary.json").read_text()) if summed.returncode == 0 else {}
|
| 121 |
+
status = 0 if not any(codes) and result.get("status") == "complete" else 1
|
| 122 |
+
info.update(status="complete" if status == 0 else "failed", exit_codes=codes,
|
| 123 |
+
finished_unix=time.time())
|
| 124 |
+
save(receipt, info)
|
| 125 |
+
(root / "eval_exit_code").write_text(f"{status}\n")
|
| 126 |
+
print("EVALUATION_EXIT", status, flush=True)
|
| 127 |
+
return status
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
if __name__ == "__main__":
|
| 131 |
+
raise SystemExit(main())
|
evaluations/base_pass8_16k_seed42/code/tests/test_math_eval_suite.py
ADDED
|
@@ -0,0 +1,106 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Scientific-metric boundaries: final answers, pass@8 and trial identity."""
|
| 2 |
+
import pytest
|
| 3 |
+
|
| 4 |
+
import json
|
| 5 |
+
import importlib.util
|
| 6 |
+
from pathlib import Path
|
| 7 |
+
|
| 8 |
+
from rltt_baseline.eval_math_suite import (
|
| 9 |
+
PROTOCOL, execution_parameters, grade, journal_records, last_boxed,
|
| 10 |
+
normalize_row, sample_seed, summary, validate_trial,
|
| 11 |
+
)
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def test_last_complete_box_with_intermediate_reasoning():
|
| 15 |
+
text = r"<think>First try \boxed{11}.</think>Final: \boxed{\frac{1}{2}}."
|
| 16 |
+
assert last_boxed(text) == r"\frac{1}{2}"
|
| 17 |
+
assert last_boxed(text + r" unfinished \boxed{") == r"\frac{1}{2}"
|
| 18 |
+
assert last_boxed("I tried 42 and 12 without a boxed answer") is None
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def test_math_equivalence_preserves_answer_meaning():
|
| 22 |
+
pytest.importorskip("math_verify")
|
| 23 |
+
assert grade(r"\frac{1}{2}", "0.5")
|
| 24 |
+
assert grade("001", "1")
|
| 25 |
+
assert not grade("42", "43")
|
| 26 |
+
assert not grade("50", r"50\%")
|
| 27 |
+
assert not grade("0", None)
|
| 28 |
+
assert grade("-2, 2", "2, -2")
|
| 29 |
+
assert not grade("-2, 2", "2")
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def test_all_olympiad_answer_components_are_kept():
|
| 33 |
+
r = normalize_row("olympiadbench", {"id": 1, "question": "Find all roots",
|
| 34 |
+
"final_answer": ["$-2$", "$2$"]}, 0)
|
| 35 |
+
assert r["gold"] == "-2, 2"
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def test_seeds_are_reproducible_and_distinct():
|
| 39 |
+
a = [sample_seed("aime24/1", i) for i in range(8)]
|
| 40 |
+
assert len(set(a)) == 8
|
| 41 |
+
assert a == [sample_seed("aime24/1", i) for i in range(8)]
|
| 42 |
+
assert a != [sample_seed("aime25/1", i) for i in range(8)]
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def test_pass8_counts_any_correct_only_after_eight_samples():
|
| 46 |
+
rows = [{"uid": "aime24/1", "dataset": "aime24"},
|
| 47 |
+
{"uid": "aime24/2", "dataset": "aime24"}]
|
| 48 |
+
records = [{"uid": "aime24/1", "replicate": i, "correct": i == 7} for i in range(8)]
|
| 49 |
+
partial = summary(rows, records)
|
| 50 |
+
assert partial["status"] == "partial"
|
| 51 |
+
assert partial["datasets"]["aime24"]["completed_questions"] == 1
|
| 52 |
+
assert partial["datasets"]["aime24"]["pass_at_8"] == 1
|
| 53 |
+
records += [{"uid": "aime24/2", "replicate": i, "correct": False} for i in range(8)]
|
| 54 |
+
final = summary(rows, records)
|
| 55 |
+
assert final["status"] == "complete"
|
| 56 |
+
assert final["datasets"]["aime24"]["pass_at_8"] == 0.5
|
| 57 |
+
assert final["datasets"]["aime24"]["mean_sample_accuracy"] == 1 / 16
|
| 58 |
+
with pytest.raises(ValueError, match="Duplicate"):
|
| 59 |
+
summary(rows, records + [records[0]])
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def test_runtime_concurrency_is_independent_of_pass8_and_output_budget():
|
| 63 |
+
p = {"concurrency_per_gpu": 16, "max_num_batched_tokens": 32768,
|
| 64 |
+
"gpu_memory_utilization": 0.85, "k": 8, "max_new_tokens": 16384}
|
| 65 |
+
assert execution_parameters(p) == (16, 32768, 0.85)
|
| 66 |
+
assert p["k"] == 8 and p["max_new_tokens"] == 16384
|
| 67 |
+
with pytest.raises(ValueError):
|
| 68 |
+
execution_parameters({**p, "max_num_batched_tokens": 8})
|
| 69 |
+
with pytest.raises(ValueError):
|
| 70 |
+
execution_parameters({**p, "gpu_memory_utilization": 1.1})
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def test_incomplete_journal_tail_is_never_counted_or_appended(tmp_path):
|
| 74 |
+
path = tmp_path / "rank0.jsonl"
|
| 75 |
+
path.write_bytes(b'{"uid":"done"}\n{"uid":"unfinished"')
|
| 76 |
+
assert list(journal_records(path, allow_incomplete=True)) == [{"uid": "done"}]
|
| 77 |
+
with pytest.raises(ValueError, match="Incomplete"):
|
| 78 |
+
list(journal_records(path))
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def test_resume_preserves_committed_bytes_seeds_and_sharding(tmp_path):
|
| 82 |
+
path = Path(__file__).resolve().parents[1] / "scripts/resume_math_eval.py"
|
| 83 |
+
spec = importlib.util.spec_from_file_location("resume_math_eval", path)
|
| 84 |
+
module = importlib.util.module_from_spec(spec)
|
| 85 |
+
spec.loader.exec_module(module)
|
| 86 |
+
copy_committed = module.copy_committed
|
| 87 |
+
row = {"uid": "aime24/1", "dataset": "aime24", "gold": "42",
|
| 88 |
+
"prompt_token_ids": [1, 2]}
|
| 89 |
+
r = {"uid": row["uid"], "replicate": 0, "dataset": row["dataset"],
|
| 90 |
+
"gold": row["gold"], "prompt_tokens": 2, "protocol": PROTOCOL,
|
| 91 |
+
"seed": sample_seed(row["uid"], 0), "correct": True,
|
| 92 |
+
"generated_tokens": 1, "token_ids": [42], "rank": 0}
|
| 93 |
+
known = {row["uid"]: row}
|
| 94 |
+
line = (json.dumps(r) + "\n").encode()
|
| 95 |
+
source = tmp_path / "source.jsonl"
|
| 96 |
+
source.write_bytes(line + b'{"interrupted":')
|
| 97 |
+
target = tmp_path / "target.jsonl"
|
| 98 |
+
info = copy_committed(source, target, known, 0)
|
| 99 |
+
assert target.read_bytes() == line
|
| 100 |
+
assert info["responses"] == 1 and info["incomplete_tail_bytes"] > 0
|
| 101 |
+
assert validate_trial(r, known) == (row["uid"], 0)
|
| 102 |
+
with pytest.raises(ValueError, match="frozen"):
|
| 103 |
+
validate_trial({**r, "seed": r["seed"] + 1}, known)
|
| 104 |
+
source.write_bytes(line + line)
|
| 105 |
+
with pytest.raises(ValueError, match="Duplicate"):
|
| 106 |
+
copy_committed(source, tmp_path / "duplicate.jsonl", known, 0)
|
evaluations/base_pass8_16k_seed42/data/questions.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/data/sources.json
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"aime24": {
|
| 3 |
+
"name": "aime24",
|
| 4 |
+
"repo": "math-ai/aime24",
|
| 5 |
+
"revision": "83a7f387baaa524a8bda0022eac0541582297103",
|
| 6 |
+
"path_in_repo": "test-00000-of-00001.parquet",
|
| 7 |
+
"url": "https://huggingface.co/datasets/math-ai/aime24/resolve/83a7f387baaa524a8bda0022eac0541582297103/test-00000-of-00001.parquet",
|
| 8 |
+
"raw_sha256": "2602156eb2a0ab64393dc46617a3bfee3451ae3c985a7ea6ac2b55436e094a38",
|
| 9 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/aime24.parquet",
|
| 10 |
+
"rows": 30
|
| 11 |
+
},
|
| 12 |
+
"aime25": {
|
| 13 |
+
"name": "aime25",
|
| 14 |
+
"repo": "math-ai/aime25",
|
| 15 |
+
"revision": "563bb8404243c5f09de6ec262f2db674fe5bce9b",
|
| 16 |
+
"path_in_repo": "test.jsonl",
|
| 17 |
+
"url": "https://huggingface.co/datasets/math-ai/aime25/resolve/563bb8404243c5f09de6ec262f2db674fe5bce9b/test.jsonl",
|
| 18 |
+
"raw_sha256": "b4e273c02d3e7fe1b74b59eae768fc8230bfb0f79539890cb56f4361caac0331",
|
| 19 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/aime25.jsonl",
|
| 20 |
+
"rows": 30
|
| 21 |
+
},
|
| 22 |
+
"aime26": {
|
| 23 |
+
"name": "aime26",
|
| 24 |
+
"repo": "math-ai/aime26",
|
| 25 |
+
"revision": "79037aebdb6580008fb960d17cb21fd3099083e3",
|
| 26 |
+
"path_in_repo": "aime2026.jsonl",
|
| 27 |
+
"url": "https://huggingface.co/datasets/math-ai/aime26/resolve/79037aebdb6580008fb960d17cb21fd3099083e3/aime2026.jsonl",
|
| 28 |
+
"raw_sha256": "52822957957a3f577d1e9706c36a66a8108a3f99b6aff424cfb72dff0094a9ee",
|
| 29 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/aime26.jsonl",
|
| 30 |
+
"rows": 30
|
| 31 |
+
},
|
| 32 |
+
"amc23": {
|
| 33 |
+
"name": "amc23",
|
| 34 |
+
"repo": "QwenLM/Qwen2.5-Math",
|
| 35 |
+
"revision": "a45202bd16f1ec06f433442dc1152d0074773465",
|
| 36 |
+
"path_in_repo": "evaluation/data/amc23/test.jsonl",
|
| 37 |
+
"url": "https://raw.githubusercontent.com/QwenLM/Qwen2.5-Math/a45202bd16f1ec06f433442dc1152d0074773465/evaluation/data/amc23/test.jsonl",
|
| 38 |
+
"raw_sha256": "43707e1ca784602d8479de24ee8ce36609dbebc46dd152b0b12a46611aa898a4",
|
| 39 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/amc23.jsonl",
|
| 40 |
+
"rows": 40
|
| 41 |
+
},
|
| 42 |
+
"olympiadbench": {
|
| 43 |
+
"name": "olympiadbench",
|
| 44 |
+
"repo": "QwenLM/Qwen2.5-Math",
|
| 45 |
+
"revision": "a45202bd16f1ec06f433442dc1152d0074773465",
|
| 46 |
+
"path_in_repo": "evaluation/data/olympiadbench/test.jsonl",
|
| 47 |
+
"url": "https://raw.githubusercontent.com/QwenLM/Qwen2.5-Math/a45202bd16f1ec06f433442dc1152d0074773465/evaluation/data/olympiadbench/test.jsonl",
|
| 48 |
+
"raw_sha256": "bb8c1e38a16ea9eacb77530315a674d883f267fbff79183effc53681b4fcc416",
|
| 49 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/olympiadbench.jsonl",
|
| 50 |
+
"rows": 675
|
| 51 |
+
},
|
| 52 |
+
"minerva_math": {
|
| 53 |
+
"name": "minerva_math",
|
| 54 |
+
"repo": "QwenLM/Qwen2.5-Math",
|
| 55 |
+
"revision": "a45202bd16f1ec06f433442dc1152d0074773465",
|
| 56 |
+
"path_in_repo": "evaluation/data/minerva_math/test.jsonl",
|
| 57 |
+
"url": "https://raw.githubusercontent.com/QwenLM/Qwen2.5-Math/a45202bd16f1ec06f433442dc1152d0074773465/evaluation/data/minerva_math/test.jsonl",
|
| 58 |
+
"raw_sha256": "0e656b430d3a0a1dab6aa853f453ba3805e28e2cabe4fc72486a897cc09d8ba2",
|
| 59 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/minerva_math.jsonl",
|
| 60 |
+
"rows": 272
|
| 61 |
+
},
|
| 62 |
+
"hmmt_feb25": {
|
| 63 |
+
"name": "hmmt_feb25",
|
| 64 |
+
"repo": "MathArena/hmmt_feb_2025",
|
| 65 |
+
"revision": "6fdc4277120810ff75aa22d2d5489b91f7a262a1",
|
| 66 |
+
"path_in_repo": "data/train-00000-of-00001.parquet",
|
| 67 |
+
"url": "https://huggingface.co/datasets/MathArena/hmmt_feb_2025/resolve/6fdc4277120810ff75aa22d2d5489b91f7a262a1/data/train-00000-of-00001.parquet",
|
| 68 |
+
"raw_sha256": "ccca64d248b89dc0765e40b893fe4ebeab242226a4336027d95d256886a3d5f2",
|
| 69 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/hmmt_feb25.parquet",
|
| 70 |
+
"rows": 30
|
| 71 |
+
},
|
| 72 |
+
"hmmt_nov25": {
|
| 73 |
+
"name": "hmmt_nov25",
|
| 74 |
+
"repo": "MathArena/hmmt_nov_2025",
|
| 75 |
+
"revision": "118dbfb45c4c9467c672268ed55166642897aa46",
|
| 76 |
+
"path_in_repo": "data/train-00000-of-00001.parquet",
|
| 77 |
+
"url": "https://huggingface.co/datasets/MathArena/hmmt_nov_2025/resolve/118dbfb45c4c9467c672268ed55166642897aa46/data/train-00000-of-00001.parquet",
|
| 78 |
+
"raw_sha256": "87ef3faaae0367b51bfa966d948106a507f39058a197e311f74fa5a65c8fe552",
|
| 79 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/hmmt_nov25.parquet",
|
| 80 |
+
"rows": 30
|
| 81 |
+
},
|
| 82 |
+
"math500": {
|
| 83 |
+
"name": "math500",
|
| 84 |
+
"repo": "HuggingFaceH4/MATH-500",
|
| 85 |
+
"revision": "6e4ed1a2a79af7d8630a6b768ec859cb5af4d3be",
|
| 86 |
+
"path_in_repo": "test.jsonl",
|
| 87 |
+
"raw_sha256": "35dc41080a3680858b27fa7e0533d2d547825316fc5dafe5d316f4ccc5a06132",
|
| 88 |
+
"raw_file": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1/data/raw/math500.jsonl",
|
| 89 |
+
"rows": 500
|
| 90 |
+
}
|
| 91 |
+
}
|
evaluations/base_pass8_16k_seed42/eval_exit_code
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
0
|
evaluations/base_pass8_16k_seed42/logs/coordinator.log
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
WORKER_SUBMITTED rank=0 gpu=1 pid=3100957
|
| 2 |
+
WORKER_SUBMITTED rank=1 gpu=2 pid=3101276
|
| 3 |
+
WORKER_SUBMITTED rank=2 gpu=3 pid=3101277
|
| 4 |
+
WORKER_SUBMITTED rank=3 gpu=4 pid=3101278
|
| 5 |
+
WORKER_SUBMITTED rank=4 gpu=5 pid=3101279
|
| 6 |
+
WORKER_SUBMITTED rank=5 gpu=6 pid=3101283
|
| 7 |
+
WORKER_SUBMITTED rank=6 gpu=7 pid=3101287
|
| 8 |
+
SUMMARY {"protocol": "looprl-math-pass8-16k-v1", "responses_completed": 13096, "responses_expected": 13096, "status": "complete", "datasets": {"aime24": {"questions": 30, "completed_questions": 30, "pass_at_8": 0.6333333333333333, "mean_sample_accuracy": 0.45416666666666666, "status": "complete"}, "aime25": {"questions": 30, "completed_questions": 30, "pass_at_8": 0.6, "mean_sample_accuracy": 0.4083333333333333, "status": "complete"}, "aime26": {"questions": 30, "completed_questions": 30, "pass_at_8": 0.7, "mean_sample_accuracy": 0.42083333333333334, "status": "complete"}, "math500": {"questions": 500, "completed_questions": 500, "pass_at_8": 0.974, "mean_sample_accuracy": 0.91775, "status": "complete"}, "amc23": {"questions": 40, "completed_questions": 40, "pass_at_8": 0.925, "mean_sample_accuracy": 0.85625, "status": "complete"}, "olympiadbench": {"questions": 675, "completed_questions": 675, "pass_at_8": 0.7407407407407407, "mean_sample_accuracy": 0.5894444444444444, "status": "complete"}, "minerva_math": {"questions": 272, "completed_questions": 272, "pass_at_8": 0.5441176470588235, "mean_sample_accuracy": 0.4071691176470588, "status": "complete"}, "hmmt_feb25": {"questions": 30, "completed_questions": 30, "pass_at_8": 0.4, "mean_sample_accuracy": 0.24166666666666667, "status": "complete"}, "hmmt_nov25": {"questions": 30, "completed_questions": 30, "pass_at_8": 0.5666666666666667, "mean_sample_accuracy": 0.2375, "status": "complete"}}}
|
| 9 |
+
EVALUATION_EXIT 0
|
evaluations/base_pass8_16k_seed42/logs/rank0.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank1.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank2.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank3.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank4.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank5.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/logs/rank6.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
evaluations/base_pass8_16k_seed42/preflight/code_hashes.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rltt_baseline/eval_math_suite.py": "617694922ad4be85938cd17509a6f839e81d702d6d228dc4a495d88f2ab62dd0",
|
| 3 |
+
"scripts/run_math_eval_suite.py": "c2b6bd715b2a59aed546e2189edd9731ac065208853e63644da7f20910c169e1",
|
| 4 |
+
"scripts/resume_math_eval.py": "6b477741650c722c1b9a38b538ae9b97291a98d0eb2a5fbcf9048765c76feef4",
|
| 5 |
+
"tests/test_math_eval_suite.py": "657e9b96e78675c49d91c68d8055a6b2b421a6e5cf3b2047e8e17cebd0495f23"
|
| 6 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank0.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 0,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3100957,
|
| 5 |
+
"cuda_visible_devices": "1",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 1079,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank1.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 1,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101276,
|
| 5 |
+
"cuda_visible_devices": "2",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 964,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank2.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 2,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101277,
|
| 5 |
+
"cuda_visible_devices": "3",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 979,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank3.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 3,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101278,
|
| 5 |
+
"cuda_visible_devices": "4",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 934,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank4.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 4,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101279,
|
| 5 |
+
"cuda_visible_devices": "5",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 951,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank5.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 5,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101283,
|
| 5 |
+
"cuda_visible_devices": "6",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 1062,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/engine_rank6.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 6,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101287,
|
| 5 |
+
"cuda_visible_devices": "7",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 978,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2"
|
| 32 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/launch_receipt.json
ADDED
|
@@ -0,0 +1,78 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"logical_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 3 |
+
"pid": 3100951,
|
| 4 |
+
"gpus": [
|
| 5 |
+
1,
|
| 6 |
+
2,
|
| 7 |
+
3,
|
| 8 |
+
4,
|
| 9 |
+
5,
|
| 10 |
+
6,
|
| 11 |
+
7
|
| 12 |
+
],
|
| 13 |
+
"workers": 7,
|
| 14 |
+
"concurrency_per_gpu": 16,
|
| 15 |
+
"global_concurrency_limit": 112,
|
| 16 |
+
"protocol_sha256": "253ba69f5896ec554c0fb9ad0762e2ab18df19299c0fc2615d8989b7508fa285",
|
| 17 |
+
"started_unix": 1793619807.019661,
|
| 18 |
+
"status": "complete",
|
| 19 |
+
"children": [
|
| 20 |
+
{
|
| 21 |
+
"rank": 0,
|
| 22 |
+
"gpu": 1,
|
| 23 |
+
"pid": 3100957,
|
| 24 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank0.log"
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"rank": 1,
|
| 28 |
+
"gpu": 2,
|
| 29 |
+
"pid": 3101276,
|
| 30 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank1.log"
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"rank": 2,
|
| 34 |
+
"gpu": 3,
|
| 35 |
+
"pid": 3101277,
|
| 36 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank2.log"
|
| 37 |
+
},
|
| 38 |
+
{
|
| 39 |
+
"rank": 3,
|
| 40 |
+
"gpu": 4,
|
| 41 |
+
"pid": 3101278,
|
| 42 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank3.log"
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"rank": 4,
|
| 46 |
+
"gpu": 5,
|
| 47 |
+
"pid": 3101279,
|
| 48 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank4.log"
|
| 49 |
+
},
|
| 50 |
+
{
|
| 51 |
+
"rank": 5,
|
| 52 |
+
"gpu": 6,
|
| 53 |
+
"pid": 3101283,
|
| 54 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank5.log"
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"rank": 6,
|
| 58 |
+
"gpu": 7,
|
| 59 |
+
"pid": 3101287,
|
| 60 |
+
"log": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2/logs/rank6.log"
|
| 61 |
+
}
|
| 62 |
+
],
|
| 63 |
+
"first_worker_verified": {
|
| 64 |
+
"responses_resumed": 1079,
|
| 65 |
+
"responses_completed": 1091,
|
| 66 |
+
"kv_capacity_tokens": 304688
|
| 67 |
+
},
|
| 68 |
+
"exit_codes": [
|
| 69 |
+
0,
|
| 70 |
+
0,
|
| 71 |
+
0,
|
| 72 |
+
0,
|
| 73 |
+
0,
|
| 74 |
+
0,
|
| 75 |
+
0
|
| 76 |
+
],
|
| 77 |
+
"finished_unix": 1793626974.673925
|
| 78 |
+
}
|
evaluations/base_pass8_16k_seed42/preflight/resume.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"parent_root": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1",
|
| 3 |
+
"new_root": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 4 |
+
"source_protocol_sha256": "cb56846c4b8ada93834abc9a0bada8bfa92be1187656e4ee96d151936d6386a1",
|
| 5 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26",
|
| 6 |
+
"model_weights_sha256": "6730581fd366f6d1689728c61637520228385cdfc17c0b24f19aa840b00565fb",
|
| 7 |
+
"resumed_responses": 6947,
|
| 8 |
+
"world_size": 7,
|
| 9 |
+
"gpus": [
|
| 10 |
+
1,
|
| 11 |
+
2,
|
| 12 |
+
3,
|
| 13 |
+
4,
|
| 14 |
+
5,
|
| 15 |
+
6,
|
| 16 |
+
7
|
| 17 |
+
],
|
| 18 |
+
"runtime_before": {
|
| 19 |
+
"concurrency_per_gpu": 8,
|
| 20 |
+
"max_num_batched_tokens": 4096,
|
| 21 |
+
"gpu_memory_utilization": 0.6
|
| 22 |
+
},
|
| 23 |
+
"runtime_after": {
|
| 24 |
+
"concurrency_per_gpu": 16,
|
| 25 |
+
"max_num_batched_tokens": 32768,
|
| 26 |
+
"gpu_memory_utilization": 0.85
|
| 27 |
+
},
|
| 28 |
+
"sampling_identity_unchanged": true,
|
| 29 |
+
"journals": [
|
| 30 |
+
{
|
| 31 |
+
"rank": 0,
|
| 32 |
+
"responses": 1079,
|
| 33 |
+
"bytes": 53077469,
|
| 34 |
+
"committed_sha256": "ddb35885d11d4a66772c66cd03f1f85dbd60b9244df676ae7bc884786c09db4f",
|
| 35 |
+
"incomplete_tail_bytes": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"rank": 1,
|
| 39 |
+
"responses": 964,
|
| 40 |
+
"bytes": 50878634,
|
| 41 |
+
"committed_sha256": "2abde94d8fadab0893083f9a3ecb98be9f7c35bfd6b807aa30b06ee90d99a692",
|
| 42 |
+
"incomplete_tail_bytes": 0
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"rank": 2,
|
| 46 |
+
"responses": 979,
|
| 47 |
+
"bytes": 52470037,
|
| 48 |
+
"committed_sha256": "c0d4d21e2337e94457a87c36f9778851d7fd99738a622fe0dc6aa5e81f83b0e9",
|
| 49 |
+
"incomplete_tail_bytes": 0
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"rank": 3,
|
| 53 |
+
"responses": 934,
|
| 54 |
+
"bytes": 52189995,
|
| 55 |
+
"committed_sha256": "06c97f27c2e4ee6eecc35116991db27dcb2db8c6f9f23224a944de64eac4d6d9",
|
| 56 |
+
"incomplete_tail_bytes": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"rank": 4,
|
| 60 |
+
"responses": 951,
|
| 61 |
+
"bytes": 51719738,
|
| 62 |
+
"committed_sha256": "3afe1088932fca2163312ee9c42f819b447808f9f7ea84ad1b5119e4356cc1c6",
|
| 63 |
+
"incomplete_tail_bytes": 0
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"rank": 5,
|
| 67 |
+
"responses": 1062,
|
| 68 |
+
"bytes": 52532721,
|
| 69 |
+
"committed_sha256": "edaca60916c9237f029d8fd49180b7bddde328787c0bb00e203ab7038462d62e",
|
| 70 |
+
"incomplete_tail_bytes": 0
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"rank": 6,
|
| 74 |
+
"responses": 978,
|
| 75 |
+
"bytes": 51705486,
|
| 76 |
+
"committed_sha256": "91a666243dbe3da058b7c2aceb97b2929e4d17864207b02a6d86eb3ad2a86429",
|
| 77 |
+
"incomplete_tail_bytes": 0
|
| 78 |
+
}
|
| 79 |
+
],
|
| 80 |
+
"prepared_unix": 1793619755.9565609,
|
| 81 |
+
"status": "prepared_not_started"
|
| 82 |
+
}
|
evaluations/base_pass8_16k_seed42/protocol.json
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"protocol": "looprl-math-pass8-16k-v1",
|
| 3 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 4 |
+
"model_id": "ByteDance/Ouro-1.4B-Thinking",
|
| 5 |
+
"revision": "3aaa2224253a92ca45cf2e3d427c360e1ef9c93d",
|
| 6 |
+
"weights_sha256": "6730581fd366f6d1689728c61637520228385cdfc17c0b24f19aa840b00565fb",
|
| 7 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26",
|
| 8 |
+
"datasets": {
|
| 9 |
+
"aime24": 30,
|
| 10 |
+
"aime25": 30,
|
| 11 |
+
"aime26": 30,
|
| 12 |
+
"math500": 500,
|
| 13 |
+
"amc23": 40,
|
| 14 |
+
"olympiadbench": 675,
|
| 15 |
+
"minerva_math": 272,
|
| 16 |
+
"hmmt_feb25": 30,
|
| 17 |
+
"hmmt_nov25": 30
|
| 18 |
+
},
|
| 19 |
+
"questions": 1637,
|
| 20 |
+
"responses": 13096,
|
| 21 |
+
"k": 8,
|
| 22 |
+
"max_new_tokens": 16384,
|
| 23 |
+
"max_prompt_tokens": 1342,
|
| 24 |
+
"max_model_len": 17920,
|
| 25 |
+
"loops": 4,
|
| 26 |
+
"dtype": "bfloat16",
|
| 27 |
+
"temperature": 0.6,
|
| 28 |
+
"top_p": 0.95,
|
| 29 |
+
"top_k": -1,
|
| 30 |
+
"concurrency_per_gpu": 16,
|
| 31 |
+
"max_num_batched_tokens": 32768,
|
| 32 |
+
"gpu_memory_utilization": 0.85,
|
| 33 |
+
"tensor_parallel_size": 1,
|
| 34 |
+
"prompt": "zero-shot, Thinking chat template, training math instruction",
|
| 35 |
+
"grader": "last-complete-box-plain-or-math-verify-v1",
|
| 36 |
+
"length_stops": "recorded; scored by the same last-box rule",
|
| 37 |
+
"source_sha256": "617694922ad4be85938cd17509a6f839e81d702d6d228dc4a495d88f2ab62dd0",
|
| 38 |
+
"versions": {
|
| 39 |
+
"vllm": "0.12.0",
|
| 40 |
+
"torch": "2.9.0",
|
| 41 |
+
"transformers": "4.57.3",
|
| 42 |
+
"math-verify": "0.8.0"
|
| 43 |
+
},
|
| 44 |
+
"parent_root": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/zlh/looprl/eval_base_1p4b_9bench_pass8_16k_20261102_r1",
|
| 45 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 46 |
+
"responses_resumed": 6947
|
| 47 |
+
}
|
evaluations/base_pass8_16k_seed42/provenance/parent_protocol.json
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"protocol": "looprl-math-pass8-16k-v1",
|
| 3 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 4 |
+
"model_id": "ByteDance/Ouro-1.4B-Thinking",
|
| 5 |
+
"revision": "3aaa2224253a92ca45cf2e3d427c360e1ef9c93d",
|
| 6 |
+
"weights_sha256": "6730581fd366f6d1689728c61637520228385cdfc17c0b24f19aa840b00565fb",
|
| 7 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26",
|
| 8 |
+
"datasets": {
|
| 9 |
+
"aime24": 30,
|
| 10 |
+
"aime25": 30,
|
| 11 |
+
"aime26": 30,
|
| 12 |
+
"math500": 500,
|
| 13 |
+
"amc23": 40,
|
| 14 |
+
"olympiadbench": 675,
|
| 15 |
+
"minerva_math": 272,
|
| 16 |
+
"hmmt_feb25": 30,
|
| 17 |
+
"hmmt_nov25": 30
|
| 18 |
+
},
|
| 19 |
+
"questions": 1637,
|
| 20 |
+
"responses": 13096,
|
| 21 |
+
"k": 8,
|
| 22 |
+
"max_new_tokens": 16384,
|
| 23 |
+
"max_prompt_tokens": 1342,
|
| 24 |
+
"max_model_len": 17920,
|
| 25 |
+
"loops": 4,
|
| 26 |
+
"dtype": "bfloat16",
|
| 27 |
+
"temperature": 0.6,
|
| 28 |
+
"top_p": 0.95,
|
| 29 |
+
"top_k": -1,
|
| 30 |
+
"concurrency_per_gpu": 8,
|
| 31 |
+
"max_num_batched_tokens": 4096,
|
| 32 |
+
"gpu_memory_utilization": 0.6,
|
| 33 |
+
"tensor_parallel_size": 1,
|
| 34 |
+
"prompt": "zero-shot, Thinking chat template, training math instruction",
|
| 35 |
+
"grader": "last-complete-box-plain-or-math-verify-v1",
|
| 36 |
+
"length_stops": "recorded; scored by the same last-box rule",
|
| 37 |
+
"source_sha256": "f2894ca2047c4f20486fcbdaef0451fa6e9c575a639a3306eff55d95b8e4a435",
|
| 38 |
+
"versions": {
|
| 39 |
+
"vllm": "0.12.0",
|
| 40 |
+
"torch": "2.9.0",
|
| 41 |
+
"transformers": "4.57.3",
|
| 42 |
+
"math-verify": "0.8.0"
|
| 43 |
+
}
|
| 44 |
+
}
|
evaluations/base_pass8_16k_seed42/responses/rank0.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b474e6b519151dfee60ae1ed25cc884e73132a0f40b85d7648568fc5cc120880
|
| 3 |
+
size 82830095
|
evaluations/base_pass8_16k_seed42/responses/rank1.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b74eb39bc13e9863c5e00130ab8c1859f5530c6bb487ab0f58cf88143bcc0d60
|
| 3 |
+
size 87703153
|
evaluations/base_pass8_16k_seed42/responses/rank2.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d6ba8f53116beca816cca835c8107e2e5a1734a7f2188fd1ba91a8cb733d7683
|
| 3 |
+
size 91204987
|
evaluations/base_pass8_16k_seed42/responses/rank3.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:31bc6e2a07cdaf5725d4a6b404b905ee55cb75b6abc78d388107826526ee7ec3
|
| 3 |
+
size 88851916
|
evaluations/base_pass8_16k_seed42/responses/rank4.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9c276b840782c58b15b43491fae031f789d255fc7156739e836db7dd979a6de0
|
| 3 |
+
size 92933776
|
evaluations/base_pass8_16k_seed42/responses/rank5.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ffe3fb47b2c7b3baeae904a432c19ccea5b3c26939578e78197636f89ed891a6
|
| 3 |
+
size 85207799
|
evaluations/base_pass8_16k_seed42/responses/rank6.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:44bc336b34dae4d1352a74cfb331f4a6edf204517247880a1e4d212097dad2b8
|
| 3 |
+
size 90354521
|
evaluations/base_pass8_16k_seed42/summary.json
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"protocol": "looprl-math-pass8-16k-v1",
|
| 3 |
+
"responses_completed": 13096,
|
| 4 |
+
"responses_expected": 13096,
|
| 5 |
+
"status": "complete",
|
| 6 |
+
"datasets": {
|
| 7 |
+
"aime24": {
|
| 8 |
+
"questions": 30,
|
| 9 |
+
"completed_questions": 30,
|
| 10 |
+
"pass_at_8": 0.6333333333333333,
|
| 11 |
+
"mean_sample_accuracy": 0.45416666666666666,
|
| 12 |
+
"status": "complete"
|
| 13 |
+
},
|
| 14 |
+
"aime25": {
|
| 15 |
+
"questions": 30,
|
| 16 |
+
"completed_questions": 30,
|
| 17 |
+
"pass_at_8": 0.6,
|
| 18 |
+
"mean_sample_accuracy": 0.4083333333333333,
|
| 19 |
+
"status": "complete"
|
| 20 |
+
},
|
| 21 |
+
"aime26": {
|
| 22 |
+
"questions": 30,
|
| 23 |
+
"completed_questions": 30,
|
| 24 |
+
"pass_at_8": 0.7,
|
| 25 |
+
"mean_sample_accuracy": 0.42083333333333334,
|
| 26 |
+
"status": "complete"
|
| 27 |
+
},
|
| 28 |
+
"math500": {
|
| 29 |
+
"questions": 500,
|
| 30 |
+
"completed_questions": 500,
|
| 31 |
+
"pass_at_8": 0.974,
|
| 32 |
+
"mean_sample_accuracy": 0.91775,
|
| 33 |
+
"status": "complete"
|
| 34 |
+
},
|
| 35 |
+
"amc23": {
|
| 36 |
+
"questions": 40,
|
| 37 |
+
"completed_questions": 40,
|
| 38 |
+
"pass_at_8": 0.925,
|
| 39 |
+
"mean_sample_accuracy": 0.85625,
|
| 40 |
+
"status": "complete"
|
| 41 |
+
},
|
| 42 |
+
"olympiadbench": {
|
| 43 |
+
"questions": 675,
|
| 44 |
+
"completed_questions": 675,
|
| 45 |
+
"pass_at_8": 0.7407407407407407,
|
| 46 |
+
"mean_sample_accuracy": 0.5894444444444444,
|
| 47 |
+
"status": "complete"
|
| 48 |
+
},
|
| 49 |
+
"minerva_math": {
|
| 50 |
+
"questions": 272,
|
| 51 |
+
"completed_questions": 272,
|
| 52 |
+
"pass_at_8": 0.5441176470588235,
|
| 53 |
+
"mean_sample_accuracy": 0.4071691176470588,
|
| 54 |
+
"status": "complete"
|
| 55 |
+
},
|
| 56 |
+
"hmmt_feb25": {
|
| 57 |
+
"questions": 30,
|
| 58 |
+
"completed_questions": 30,
|
| 59 |
+
"pass_at_8": 0.4,
|
| 60 |
+
"mean_sample_accuracy": 0.24166666666666667,
|
| 61 |
+
"status": "complete"
|
| 62 |
+
},
|
| 63 |
+
"hmmt_nov25": {
|
| 64 |
+
"questions": 30,
|
| 65 |
+
"completed_questions": 30,
|
| 66 |
+
"pass_at_8": 0.5666666666666667,
|
| 67 |
+
"mean_sample_accuracy": 0.2375,
|
| 68 |
+
"status": "complete"
|
| 69 |
+
}
|
| 70 |
+
}
|
| 71 |
+
}
|
evaluations/base_pass8_16k_seed42/validation.json
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"status": "complete_verified",
|
| 3 |
+
"questions": 1637,
|
| 4 |
+
"responses": 13096,
|
| 5 |
+
"strict_journals": true,
|
| 6 |
+
"unique_complete_replicates": true,
|
| 7 |
+
"summary_recomputed_equal": true,
|
| 8 |
+
"worker_exit_codes": [
|
| 9 |
+
0,
|
| 10 |
+
0,
|
| 11 |
+
0,
|
| 12 |
+
0,
|
| 13 |
+
0,
|
| 14 |
+
0,
|
| 15 |
+
0
|
| 16 |
+
],
|
| 17 |
+
"coordinator_exit_code": 0,
|
| 18 |
+
"results": {
|
| 19 |
+
"aime24": {
|
| 20 |
+
"correct_questions": 19,
|
| 21 |
+
"questions": 30,
|
| 22 |
+
"pass_at_8": 0.6333333333333333
|
| 23 |
+
},
|
| 24 |
+
"aime25": {
|
| 25 |
+
"correct_questions": 18,
|
| 26 |
+
"questions": 30,
|
| 27 |
+
"pass_at_8": 0.6
|
| 28 |
+
},
|
| 29 |
+
"aime26": {
|
| 30 |
+
"correct_questions": 21,
|
| 31 |
+
"questions": 30,
|
| 32 |
+
"pass_at_8": 0.7
|
| 33 |
+
},
|
| 34 |
+
"math500": {
|
| 35 |
+
"correct_questions": 487,
|
| 36 |
+
"questions": 500,
|
| 37 |
+
"pass_at_8": 0.974
|
| 38 |
+
},
|
| 39 |
+
"amc23": {
|
| 40 |
+
"correct_questions": 37,
|
| 41 |
+
"questions": 40,
|
| 42 |
+
"pass_at_8": 0.925
|
| 43 |
+
},
|
| 44 |
+
"olympiadbench": {
|
| 45 |
+
"correct_questions": 500,
|
| 46 |
+
"questions": 675,
|
| 47 |
+
"pass_at_8": 0.7407407407407407
|
| 48 |
+
},
|
| 49 |
+
"minerva_math": {
|
| 50 |
+
"correct_questions": 148,
|
| 51 |
+
"questions": 272,
|
| 52 |
+
"pass_at_8": 0.5441176470588235
|
| 53 |
+
},
|
| 54 |
+
"hmmt_feb25": {
|
| 55 |
+
"correct_questions": 12,
|
| 56 |
+
"questions": 30,
|
| 57 |
+
"pass_at_8": 0.4
|
| 58 |
+
},
|
| 59 |
+
"hmmt_nov25": {
|
| 60 |
+
"correct_questions": 17,
|
| 61 |
+
"questions": 30,
|
| 62 |
+
"pass_at_8": 0.5666666666666667
|
| 63 |
+
}
|
| 64 |
+
},
|
| 65 |
+
"journals": {
|
| 66 |
+
"rank0.jsonl": {
|
| 67 |
+
"responses": 1872,
|
| 68 |
+
"bytes": 82830095,
|
| 69 |
+
"sha256": "b474e6b519151dfee60ae1ed25cc884e73132a0f40b85d7648568fc5cc120880",
|
| 70 |
+
"git_blob_sha1": "116337724f55a34afcb5ec87ba12c73af9111db8"
|
| 71 |
+
},
|
| 72 |
+
"rank1.jsonl": {
|
| 73 |
+
"responses": 1872,
|
| 74 |
+
"bytes": 87703153,
|
| 75 |
+
"sha256": "b74eb39bc13e9863c5e00130ab8c1859f5530c6bb487ab0f58cf88143bcc0d60",
|
| 76 |
+
"git_blob_sha1": "e4492198f4fbeffbb5792380b7c913b3f03d01cd"
|
| 77 |
+
},
|
| 78 |
+
"rank2.jsonl": {
|
| 79 |
+
"responses": 1872,
|
| 80 |
+
"bytes": 91204987,
|
| 81 |
+
"sha256": "d6ba8f53116beca816cca835c8107e2e5a1734a7f2188fd1ba91a8cb733d7683",
|
| 82 |
+
"git_blob_sha1": "d5d40a6db9dc139a6d09861d2d6be998192f84cc"
|
| 83 |
+
},
|
| 84 |
+
"rank3.jsonl": {
|
| 85 |
+
"responses": 1872,
|
| 86 |
+
"bytes": 88851916,
|
| 87 |
+
"sha256": "31bc6e2a07cdaf5725d4a6b404b905ee55cb75b6abc78d388107826526ee7ec3",
|
| 88 |
+
"git_blob_sha1": "358a8ecef0d13d8d62cf16c9e89e0bbcc815e88a"
|
| 89 |
+
},
|
| 90 |
+
"rank4.jsonl": {
|
| 91 |
+
"responses": 1872,
|
| 92 |
+
"bytes": 92933776,
|
| 93 |
+
"sha256": "9c276b840782c58b15b43491fae031f789d255fc7156739e836db7dd979a6de0",
|
| 94 |
+
"git_blob_sha1": "d3c0eb742d21a455b1a7abeb69c0801134130cbc"
|
| 95 |
+
},
|
| 96 |
+
"rank5.jsonl": {
|
| 97 |
+
"responses": 1872,
|
| 98 |
+
"bytes": 85207799,
|
| 99 |
+
"sha256": "ffe3fb47b2c7b3baeae904a432c19ccea5b3c26939578e78197636f89ed891a6",
|
| 100 |
+
"git_blob_sha1": "75a0db5c7a16a47bd4db249a57700f37f1f9415d"
|
| 101 |
+
},
|
| 102 |
+
"rank6.jsonl": {
|
| 103 |
+
"responses": 1864,
|
| 104 |
+
"bytes": 90354521,
|
| 105 |
+
"sha256": "44bc336b34dae4d1352a74cfb331f4a6edf204517247880a1e4d212097dad2b8",
|
| 106 |
+
"git_blob_sha1": "537ede571f973e2ed0be5ab6478684afcad01b57"
|
| 107 |
+
}
|
| 108 |
+
},
|
| 109 |
+
"protocol_sha256": "253ba69f5896ec554c0fb9ad0762e2ab18df19299c0fc2615d8989b7508fa285",
|
| 110 |
+
"questions_sha256": "65f175b98d2b2b284c8706e00d341ecd4462ddfa003ac7a6e65f7cd2f6290a26"
|
| 111 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank0.complete.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"responses": 1872,
|
| 3 |
+
"expected": 1872,
|
| 4 |
+
"finished_unix": 1793624686.61283
|
| 5 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank0.exit_code
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
0
|
evaluations/base_pass8_16k_seed42/workers/rank0.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 0,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3100957,
|
| 5 |
+
"cuda_visible_devices": "1",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 1079,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 32 |
+
"status": "complete",
|
| 33 |
+
"active_sequences": 0,
|
| 34 |
+
"queued_sequences": 0,
|
| 35 |
+
"responses_completed": 1872,
|
| 36 |
+
"responses_expected": 1872,
|
| 37 |
+
"generated_tokens_this_session": 3945169,
|
| 38 |
+
"elapsed_seconds": 4772.316600238904,
|
| 39 |
+
"updated_unix": 1793624686.538023
|
| 40 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank1.complete.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"responses": 1872,
|
| 3 |
+
"expected": 1872,
|
| 4 |
+
"finished_unix": 1793626076.646889
|
| 5 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank1.exit_code
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
0
|
evaluations/base_pass8_16k_seed42/workers/rank1.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 1,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101276,
|
| 5 |
+
"cuda_visible_devices": "2",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 964,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 32 |
+
"status": "complete",
|
| 33 |
+
"active_sequences": 0,
|
| 34 |
+
"queued_sequences": 0,
|
| 35 |
+
"responses_completed": 1872,
|
| 36 |
+
"responses_expected": 1872,
|
| 37 |
+
"generated_tokens_this_session": 4799660,
|
| 38 |
+
"elapsed_seconds": 6028.626281596022,
|
| 39 |
+
"updated_unix": 1793626076.565179
|
| 40 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank2.complete.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"responses": 1872,
|
| 3 |
+
"expected": 1872,
|
| 4 |
+
"finished_unix": 1793626461.244524
|
| 5 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank2.exit_code
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
0
|
evaluations/base_pass8_16k_seed42/workers/rank2.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"rank": 2,
|
| 3 |
+
"world_size": 7,
|
| 4 |
+
"pid": 3101277,
|
| 5 |
+
"cuda_visible_devices": "3",
|
| 6 |
+
"engine_kwargs": {
|
| 7 |
+
"model": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 8 |
+
"tokenizer": "/mnt/tidalfs-awsatl01-ren/dataset/pai/sz/looprl-work/deepmath_sampler/scripts/deepmath_filter_portable/model_1p4b_thinking_3aaa2224",
|
| 9 |
+
"trust_remote_code": true,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"tensor_parallel_size": 1,
|
| 12 |
+
"distributed_executor_backend": "mp",
|
| 13 |
+
"hf_overrides": {
|
| 14 |
+
"total_ut_steps": 4
|
| 15 |
+
},
|
| 16 |
+
"gpu_memory_utilization": 0.85,
|
| 17 |
+
"max_model_len": 17920,
|
| 18 |
+
"max_num_seqs": 16,
|
| 19 |
+
"max_num_batched_tokens": 32768,
|
| 20 |
+
"enable_chunked_prefill": true,
|
| 21 |
+
"enable_prefix_caching": false,
|
| 22 |
+
"generation_config": "vllm",
|
| 23 |
+
"seed": 42,
|
| 24 |
+
"disable_log_stats": false,
|
| 25 |
+
"model_impl": "vllm",
|
| 26 |
+
"async_scheduling": false
|
| 27 |
+
},
|
| 28 |
+
"kv_capacity_tokens": 304688,
|
| 29 |
+
"worst_case_active_tokens": 286720,
|
| 30 |
+
"responses_resumed": 979,
|
| 31 |
+
"execution_id": "eval_base_1p4b_9bench_pass8_16k_c16_mem085_bt32768_20261102_r2",
|
| 32 |
+
"status": "complete",
|
| 33 |
+
"active_sequences": 0,
|
| 34 |
+
"queued_sequences": 0,
|
| 35 |
+
"responses_completed": 1872,
|
| 36 |
+
"responses_expected": 1872,
|
| 37 |
+
"generated_tokens_this_session": 4983412,
|
| 38 |
+
"elapsed_seconds": 6414.410004780861,
|
| 39 |
+
"updated_unix": 1793626461.16335
|
| 40 |
+
}
|
evaluations/base_pass8_16k_seed42/workers/rank3.complete.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"responses": 1872,
|
| 3 |
+
"expected": 1872,
|
| 4 |
+
"finished_unix": 1793626016.0521886
|
| 5 |
+
}
|