|
1 | 1 | # task310_qwen_all_sft_30b_full_training_s1 - Qwen all-SFT 30B full training gate |
2 | 2 |
|
3 | | -<!-- METADATA:STATUS=Open,ASSIGNEE= --> |
| 3 | +<!-- METADATA:STATUS=Blocked,ASSIGNEE=intern_nemotron_worker_5,SESSION=2 --> |
4 | 4 |
|
5 | | -## 背景 |
| 5 | +## Background |
6 | 6 |
|
7 | | -Coordinator authorized attempting a full all-SFT data->training->testing pipeline, prioritizing current 30B if resource/runtime gates remain valid. This is not a promotion claim. |
| 7 | +The user requested a full pipeline on all SFT data and evaluation on available |
| 8 | +benchmarks. The coordinator authorized an attempted gate-driven run, not a |
| 9 | +promotion claim. The previous 30B AIME attempt failed the hard non-regression |
| 10 | +rule: task301/task306 FT `14/30` was below task300 base `15/30`. |
8 | 11 |
|
9 | | -## 任务目标 |
| 12 | +## Goal |
10 | 13 |
|
11 | | -Launch full Qwen3-30B-A3B all-SFT training only after task308/task309 data gates and runtime/resource gates pass; otherwise fail closed with exact blocker. |
| 14 | +Launch full Qwen3-30B-A3B all-SFT training only after data and runtime gates |
| 15 | +pass; otherwise fail closed with exact resource/runtime/data blocker. |
12 | 16 |
|
13 | | -## 实现说明 |
| 17 | +## Scope |
14 | 18 |
|
15 | | -Use /mnt/cephfs/data/stable/models/Qwen/Qwen3-30B-A3B-Instruct-2507 unless task308/task309/task298-equivalent review blocks. Do not silently downgrade. No AIME2025 train data, task255 reuse, shared deletion, export/endpoint/promotion, main push, merge, or product-code edits. |
| 19 | +- Use current `origin/main` `172cd0e7ceaba8ad2b412d1145441dbb4c5fd122`. |
| 20 | + Lead verified this is a docs-only task310 task-doc advance from product-code |
| 21 | + baseline `ecb14173a820df377270273b9f7d9d92cb5076d2`. |
| 22 | +- Prioritize target model path: |
| 23 | + `/mnt/cephfs/data/stable/models/Qwen/Qwen3-30B-A3B-Instruct-2507`. |
| 24 | +- Before training, verify: |
| 25 | + - task308 `PASS_AUDIT`; |
| 26 | + - task309 `PASS_PACKED_CONTRACT`; |
| 27 | + - current 30B runtime/resource route is still valid or explicitly refreshed |
| 28 | + from task298/task300/task301/task306 evidence; |
| 29 | + - exact GPUs, host, parallelism, launch entrypoint, config, output roots, and |
| 30 | + checkpoint policy. |
| 31 | +- If a required gate is missing, return `BLOCK_PRETRAINING_GATE` or a more |
| 32 | + specific blocker; do not silently downgrade to 4B or a different Qwen path. |
| 33 | +- When gates pass, run full all-SFT training and record: |
| 34 | + - command/env; |
| 35 | + - model path and tokenizer; |
| 36 | + - packed root and source manifest; |
| 37 | + - LR, optimizer, train steps, schedule, seed, precision, parallelism; |
| 38 | + - loss and validation summaries; |
| 39 | + - checkpoint paths and checksum manifests; |
| 40 | + - resource usage and failure/retry notes; |
| 41 | + - handoff for task311 canary/benchmark eval. |
16 | 42 |
|
17 | | -## 验收标准 |
| 43 | +## Boundaries |
18 | 44 |
|
19 | | -- PASS_TRAINING: usable checkpoint, LR/loss/validation evidence, commands/env, resources, logs, checksums, and handoff for canary/benchmark eval are complete. |
20 | | -- BLOCK: resource/runtime/data gates fail, checkpoint is unusable, or safe 30B training cannot be launched without boundary violations. |
| 45 | +- No AIME2025 prompts or labels in training data. |
| 46 | +- No task255 reuse. |
| 47 | +- No shared deletion under `/mnt/cephfs/data/processing/lei.song`. |
| 48 | +- No export, endpoint, promotion, direct main push, merge, or product-code |
| 49 | + modification. |
| 50 | +- If export/endpoint becomes needed, task311 must treat it as eval-only and |
| 51 | + document it separately. |
21 | 52 |
|
22 | | -## 分配信息 |
| 53 | +## Expected Output |
23 | 54 |
|
24 | | -- Team:nemotron |
25 | | -- Team lead:intern_nemotron_lead |
26 | | -- Worker:intern_nemotron_worker_5 |
27 | | -- 分配方式:team_lead 创建本 task 文档后,通知 worker 接受该 task。 |
| 55 | +- Worker branch: |
| 56 | + `intern_nemotron_worker_5/task310_qwen_all_sft_30b_full_training_s1`. |
| 57 | +- Report: |
| 58 | + `workspace/tasks/task310_qwen_all_sft_30b_full_training_s1/all_sft_30b_full_training_report.md`. |
| 59 | +- Remote training root and local evidence root with launch scripts/configs, |
| 60 | + logs, loss/validation summaries, checkpoint manifests, checksums, and task311 |
| 61 | + handoff. |
| 62 | +- Mailbox report with branch/head/PR or blocker, upstream gate references, |
| 63 | + exact commands/env, GPU/resource evidence, checkpoint paths, metrics, and |
| 64 | + residual risks. |
| 65 | + |
| 66 | +## Acceptance Criteria |
| 67 | + |
| 68 | +- `PASS_TRAINING`: required gates passed, training completed with a usable |
| 69 | + Qwen3-30B-A3B all-SFT checkpoint, finite loss/validation evidence, complete |
| 70 | + commands/env/logs/checksums, and task311 handoff. |
| 71 | +- `REQUEST_CHANGES`: report/artifacts lack required config, metrics, checksums, |
| 72 | + validation, or handoff evidence. |
| 73 | +- `BLOCK`: data/runtime/resource gates are absent, launch fails without usable |
| 74 | + checkpoint, contamination risk appears, shared deletion would be needed, or |
| 75 | + the run would require unauthorized downgrade/promotion/export. |
| 76 | + |
| 77 | +## Assignment |
| 78 | + |
| 79 | +- Team: `nemotron` |
| 80 | +- Team lead: `intern_nemotron_lead` |
| 81 | +- Worker: `intern_nemotron_worker_5` |
| 82 | +- Current main: `172cd0e7ceaba8ad2b412d1145441dbb4c5fd122` |
| 83 | +- Product-code baseline: `ecb14173a820df377270273b9f7d9d92cb5076d2` |
| 84 | +- Upstream dependencies: task308, task309, prior 30B runtime/resource evidence |
| 85 | +- Downstream tasks: task311, task312 |
| 86 | +- Gate state: full training is HOLD until upstream evidence is accepted. |
| 87 | + |
| 88 | +## Current Worker Disposition |
| 89 | + |
| 90 | +Session 2 disposition: `BLOCK_PRETRAINING_GATE`. |
| 91 | + |
| 92 | +Training was not launched because the required upstream gates are not yet |
| 93 | +present as accepted evidence: |
| 94 | + |
| 95 | +- task308 has a visible worker branch |
| 96 | + `intern_nemotron_worker_1/task308_qwen_all_sft_pipeline_inventory_audit_s1` |
| 97 | + at `348cba44c02043cd6310a36ec722a68278288db2`, but no PR or |
| 98 | + `all_sft_pipeline_inventory_audit_report.md` PASS artifact is visible. |
| 99 | +- task309 has a visible worker branch |
| 100 | + `intern_nemotron_worker_2/task309_qwen_all_sft_packed_data_contract_s1` |
| 101 | + at `d054925b1792a5365738247eeb8bdec462e1e6c6`, but no PR or |
| 102 | + `all_sft_packed_data_contract_report.md` PASS artifact is visible. |
| 103 | + |
| 104 | +Report: |
| 105 | +`workspace/tasks/task310_qwen_all_sft_30b_full_training_s1/all_sft_30b_full_training_report.md`. |
0 commit comments