|
| 1 | +# task296_qwen_aime_v11_current_main_equivalence_audit_s1 - current-main equivalence audit |
| 2 | + |
| 3 | +<!-- METADATA:STATUS=ReadyForPR,ASSIGNEE=intern_nemotron_worker_1,SESSION=76 --> |
| 4 | + |
| 5 | +## Background |
| 6 | + |
| 7 | +The user requested commit/merge of the outstanding PR and then a full |
| 8 | +data/training/evaluation pipeline based on current code. Coordinator PR #312 |
| 9 | +was merged after the V11 runbook closeout: |
| 10 | + |
| 11 | +- post-#351 main: `5d8b8d850d26e785332f8b707c772d99881a1b5d` |
| 12 | +- current main after #312: `2d84ec75960fb51ba9091427638b00083625e137` |
| 13 | +- #312 mergedAt: `2026-06-02T12:13:44Z` |
| 14 | +- #312 merged head: `c7ada6134f63c88d1efcbf993452186d14ae24f3` |
| 15 | + |
| 16 | +Existing merged V11 evidence: |
| 17 | + |
| 18 | +- task285/#350 bounded Qwen3-4B SFT smoke from source head |
| 19 | + `c53095a639f0ccf8ce34afcec1bdf302cf45add6` |
| 20 | +- task293/#356 corrected AIME2025 eval from run source head |
| 21 | + `87de0a97e6c0406a4b67520faab6b11d91d9131e` |
| 22 | +- task293 metric: FT `12/30 = 0.4` versus accepted task247 Qwen3-4B base |
| 23 | + `11/30 = 0.36666666666666664` |
| 24 | +- task294/#357 independent review decision: |
| 25 | + `APPROVE_AIME_GATE_PASS_WITH_RESIDUAL` |
| 26 | +- task295/#351 runbook/provenance closeout merged at |
| 27 | + `5d8b8d850d26e785332f8b707c772d99881a1b5d` |
| 28 | + |
| 29 | +Lead preliminary read-only observation: #312 appears coordinator-docs-only, but |
| 30 | +this needs worker-owned audit evidence before lead can answer that no fresh |
| 31 | +current-main rerun is needed. |
| 32 | + |
| 33 | +## Goal |
| 34 | + |
| 35 | +Determine whether path A is valid: |
| 36 | + |
| 37 | +`A_PROVED_NO_RERUN`: task285/task293 artifacts are product-code-equivalent to |
| 38 | +current main `2d84ec75960fb51ba9091427638b00083625e137`, so no new |
| 39 | +data/training/non-AIME/AIME run is needed for the user's current-code request. |
| 40 | + |
| 41 | +If this cannot be proven, return: |
| 42 | + |
| 43 | +`B_REQUIRED_RERUN`: lead must launch a fresh current-main bounded pipeline. |
| 44 | + |
| 45 | +## Scope |
| 46 | + |
| 47 | +Read-only audit only. Use git/GitHub/artifact-manifest inspection. Do not run |
| 48 | +training, canary, AIME eval, export, endpoint, promotion, 30B, or 8-GPU. |
| 49 | + |
| 50 | +Required checks: |
| 51 | + |
| 52 | +- Fetch current origin state and verify `origin/main` is exactly |
| 53 | + `2d84ec75960fb51ba9091427638b00083625e137`. |
| 54 | +- Verify #312 merge metadata: mergedAt, mergeCommit, merged head, and file list. |
| 55 | +- Compare #312's merge delta: |
| 56 | + `git diff --name-status 5d8b8d850d26e785332f8b707c772d99881a1b5d..2d84ec75960fb51ba9091427638b00083625e137` |
| 57 | + and classify every changed file. |
| 58 | +- Confirm no #312 changed file is under product, data-prep, training, eval, |
| 59 | + harness, source, test, model, recipe, or task285/task293 artifact/script |
| 60 | + paths. |
| 61 | +- Compare task285 evidence source head |
| 62 | + `c53095a639f0ccf8ce34afcec1bdf302cf45add6` to current main for relevant |
| 63 | + training/data product paths, including at least `src/`, `recipes/`, `tests/`, |
| 64 | + `workspace/tasks/task285_qwen_aime_v11_bounded_qwen4b_sft_smoke_s1/`, and any |
| 65 | + task276/task283/task285 script/config paths referenced in the task285 report. |
| 66 | +- Compare task293 run source head |
| 67 | + `87de0a97e6c0406a4b67520faab6b11d91d9131e` to current main for relevant eval |
| 68 | + paths, including at least `src/`, `recipes/`, `tests/`, |
| 69 | + `workspace/tasks/task293_qwen_aime_v11_task285_same_harness_aime_eval_s1/`, |
| 70 | + and the task291/task293 route/evaluator scripts referenced in the report. |
| 71 | +- Explicitly verify whether |
| 72 | + `workspace/tasks/task293_qwen_aime_v11_task285_same_harness_aime_eval_s1/run_no_export_aime_eval.py` |
| 73 | + changed after task293 run source head. |
| 74 | +- Reconfirm the task285 and task293 artifact roots and key checksums from the |
| 75 | + merged reports; read-only checksum recomputation is allowed if local artifacts |
| 76 | + exist. |
| 77 | +- Carry residuals forward: task285 post-train built-in eval RC=1, task276 sparse |
| 78 | + valid/test split, task292 detokenized fallback residual, task293 |
| 79 | + `sampling_exact_parameter_match=false`. |
| 80 | + |
| 81 | +## Boundaries |
| 82 | + |
| 83 | +- Do not train, run nonzero-LR smoke, run live canary, run AIME/task243 eval, |
| 84 | + export, convert, launch an endpoint, promote, reuse task255, use AIME2025 |
| 85 | + prompts/labels as train data, delete shared files, push main, merge, run 30B, |
| 86 | + or use 8 GPUs. |
| 87 | +- Do not mutate task285/task293 artifacts. |
| 88 | +- Do not claim promotion or release clearance. |
| 89 | + |
| 90 | +## Expected Output |
| 91 | + |
| 92 | +- Worker branch: |
| 93 | + `intern_nemotron_worker_1/task296_qwen_aime_v11_current_main_equivalence_audit_s1` |
| 94 | +- PR or branch report containing |
| 95 | + `workspace/tasks/task296_qwen_aime_v11_current_main_equivalence_audit_s1/current_main_equivalence_audit_report.md` |
| 96 | +- Mailbox report to lead with: |
| 97 | + - branch/head/PR or blocker; |
| 98 | + - exact commands run; |
| 99 | + - #312 merge metadata and changed-file classification; |
| 100 | + - task285/task293 source-to-current comparisons; |
| 101 | + - artifact roots and checksums checked; |
| 102 | + - decision `A_PROVED_NO_RERUN`, `REQUEST_CHANGES`, or `B_REQUIRED_RERUN`; |
| 103 | + - residual risks and no-clearance boundary statement. |
| 104 | + |
| 105 | +## Acceptance Criteria |
| 106 | + |
| 107 | +- APPROVE: current main differs from accepted V11 evidence only by coordinator |
| 108 | + docs/provenance surfaces, relevant task285/task293 code paths are unchanged or |
| 109 | + proven artifact-equivalent, artifacts/checksums remain consistent, and all |
| 110 | + residuals/boundaries are visible. |
| 111 | +- REQUEST-CHANGES: report omits required comparisons, commands, artifact |
| 112 | + metadata, or residuals. |
| 113 | +- BLOCK: product/training/eval/data code changed materially, artifact |
| 114 | + equivalence cannot be proven, or any forbidden boundary would be required. |
| 115 | + |
| 116 | +## Assignment |
| 117 | + |
| 118 | +- Team: `nemotron` |
| 119 | +- Team lead: `intern_nemotron_lead` |
| 120 | +- Worker: `intern_nemotron_worker_1` |
| 121 | +- Related PRs: #312, #350, #351, #356, #357 |
| 122 | +- Related tasks: task247, task276, task277, task285, task291, task292, task293, |
| 123 | + task294, task295 |
0 commit comments