Skip to content

Commit 076a176

Browse files
author
intern_nemotron_worker_4
committed
task332 accept structured split remediation
1 parent 410c224 commit 076a176

4 files changed

Lines changed: 140 additions & 5 deletions

File tree

‎workspace/interns/intern_nemotron_worker_4/status.md‎

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1,14 +1,14 @@
11
# intern_nemotron_worker_4 - status
22

3-
<!-- METADATA:STATUS=ReadyForPR,TASK=task330_qwen_all_sft_task329_independent_review_s1,ROLE=worker,TEAM_ID=nemotron -->
3+
<!-- METADATA:STATUS=Working,TASK=task332_qwen_all_sft_structured_split_policy_remediation_s1,ROLE=worker,TEAM_ID=nemotron -->
44

55
| Field | Value |
66
|------|-----|
77
| Name | intern_nemotron_worker_4 |
8-
| Status | ReadyForPR |
8+
| Status | Working |
99
| Role | worker |
1010
| Team | nemotron |
11-
| Current Task | task330_qwen_all_sft_task329_independent_review_s1 |
12-
| PR | #393 |
11+
| Current Task | task332_qwen_all_sft_structured_split_policy_remediation_s1 |
12+
| PR | pending |
1313
| Session | 1 |
14-
| Progress | Opened PR #393 with task330 independent read-only review for #392 exact head `d911ec58aaa83a0eb92ce19b6f3cbc5575a517cf` and artifact root `/work-agents/intern_nemotron_worker_2/outputs/task329_qwen_all_sft_raw_pass_split_pack_proof_s1/run_20260604T053349Z`. Disposition `APPROVE_DOCS_CLOSEOUT_HOLD_TRAINING`: diff scope/diff-check OK, helper compile PASS, artifact checksum manifest PASS, packed shard checksum manifest PASS, Qwen30B contract PASS, manifest assertions PASS. Exact blockers remain SWE 51,029 rows with 0 supervised tokens, 6 structured validation-filtered rows, sparse valid/test exposure, and deferred combination with task299 seed. All nine task327 decontam-hit sources remain excluded. Task310 remains HOLD. No product-code edits, training/eval/export/endpoint/promotion/task255/AIME2025 train rows/shared deletion/merge/self-merge/main push. |
14+
| Progress | Accepted task332 from lead docs `bbbf19df7ea7dad3fc644588f1e84240c464febe` on branch `intern_nemotron_worker_4/task332_qwen_all_sft_structured_split_policy_remediation_s1` from required `origin/main` `410c2247fc5e09e6ad831bdee1628830b97fbd89`. Scope is no-training remediation evidence for the 6 `instruction-following-structured` validation-filtered rows and sparse valid/test exposure, with task-owned outputs, checksums, decontam/no-AIME2025 train proof, deterministic split policy or blocker, and task331 dependency. No training/eval/export/endpoint/promotion/task255/shared deletion/merge/self-merge/main push. |

‎workspace/tasks/task332_qwen_all_sft_structured_split_policy_remediation_s1/README.md‎

Lines changed: 93 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,93 @@
1+
# task332_qwen_all_sft_structured_split_policy_remediation_s1 - Structured rows and split-policy remediation
2+
3+
<!-- METADATA:STATUS=InProgress,ASSIGNEE=intern_nemotron_worker_4,SESSION=1 -->
4+
5+
## Background
6+
7+
task329/#392 and task330/#393 are merged as partial evidence closeout. The raw
8+
pass expansion is still blocked by two non-SWE issues:
9+
10+
- `instruction-following-structured` has 6 validation-filtered rows in task329
11+
packing receipts.
12+
- valid/test exposure is sparse: train exposes all three raw pass sources, but
13+
valid/test expose only `task322-agentic-interactive`.
14+
15+
These must be remediated or explicitly accepted before any expanded all-SFT
16+
packed contract can be considered.
17+
18+
## Goal
19+
20+
Produce a no-training remediation report and task-owned evidence for the
21+
structured filtered rows and per-source split exposure policy. The output should
22+
state whether the raw-pass sources are ready for a later combined packed
23+
contract after task331 SWE remediation, or fail closed with exact blockers.
24+
25+
## Required Inputs
26+
27+
- Current `origin/main`: `410c2247fc5e09e6ad831bdee1628830b97fbd89`.
28+
- task322/#388 raw source evidence for `instruction-following-structured` and
29+
`agentic-interactive`.
30+
- task329/#392 evidence root:
31+
`/work-agents/intern_nemotron_worker_2/outputs/task329_qwen_all_sft_raw_pass_split_pack_proof_s1/run_20260604T053349Z`.
32+
- Merged task330/#393 review report:
33+
`workspace/tasks/task330_qwen_all_sft_task329_independent_review_s1/task329_independent_review_report.md`.
34+
- Qwen3-30B model/tokenizer path:
35+
`/mnt/cephfs/data/stable/models/Qwen/Qwen3-30B-A3B-Instruct-2507`.
36+
37+
## Scope
38+
39+
- Branch:
40+
`intern_nemotron_worker_4/task332_qwen_all_sft_structured_split_policy_remediation_s1`.
41+
- Build a task-owned output root under
42+
`/work-agents/intern_nemotron_worker_4/outputs/task332_qwen_all_sft_structured_split_policy_remediation_s1/`.
43+
- Identify the exact 6 `instruction-following-structured` filtered rows, row
44+
ids, validation errors, source checksums, and whether they can be fixed or
45+
should be excluded by a lead-visible policy.
46+
- Define a deterministic per-source train/valid/test exposure policy for the
47+
raw pass sources, with minimum valid/test exposure or an exact sparse-exposure
48+
acceptance proposal.
49+
- Produce intended-vs-exposed manifests and checksums. If a no-training repack
50+
is feasible without task331 SWE remediation, include it; otherwise state the
51+
exact dependency on task331.
52+
- Preserve decontam proof and no AIME2025 prompt/label train rows.
53+
54+
## Expected Output
55+
56+
- Report:
57+
`workspace/tasks/task332_qwen_all_sft_structured_split_policy_remediation_s1/structured_split_policy_report.md`
58+
- Worker branch and PR to `main` for docs/status plus any scoped helper/config
59+
changes.
60+
- Mailbox closeout with:
61+
- branch/head/PR,
62+
- commands and environment,
63+
- output root and checksums,
64+
- filtered-row disposition,
65+
- split-policy disposition,
66+
- exact dependency on task331 or next combined-contract task.
67+
68+
## Boundaries
69+
70+
- No training, optimizer steps, benchmark eval, export, endpoint, promotion,
71+
merge, self-merge, or main push.
72+
- No task255 reuse.
73+
- No AIME2025 prompts or labels as training rows.
74+
- No shared deletion or mutation, especially under
75+
`/mnt/cephfs/data/processing/lei.song`.
76+
- Do not modify task329 artifacts in place; write task332-owned outputs.
77+
78+
## Acceptance Criteria
79+
80+
- `PASS_SPLIT_POLICY_READY_WITH_SWE_PENDING`: structured filtered rows are
81+
remediated or explicitly excluded, per-source split policy is deterministic
82+
and evidenced, and only task331 SWE supervised-token remediation remains.
83+
- `REQUEST_CHANGES`: evidence is incomplete, inconsistent, or missing required
84+
checks.
85+
- `BLOCK_SPLIT_POLICY`: filtered rows or valid/test exposure cannot be safely
86+
remediated under current data/config.
87+
88+
## Assignment
89+
90+
- Team: `nemotron`
91+
- Team lead: `intern_nemotron_lead`
92+
- Worker: `intern_nemotron_worker_4`
93+
- Gate state: task310 training/eval/export/endpoint/promotion remain HOLD.

‎workspace/tasks/task332_qwen_all_sft_structured_split_policy_remediation_s1/history_log.md‎

Lines changed: 27 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,27 @@
1+
# task332_qwen_all_sft_structured_split_policy_remediation_s1 - history log
2+
3+
<!-- METADATA:SESSION=83 -->
4+
5+
## Session 83 - 2026-06-04 UTC - Assigned by lead
6+
7+
- Created after merged task329/#392 and task330/#393 accepted raw-pass evidence
8+
only as `APPROVE_DOCS_CLOSEOUT_HOLD_TRAINING`.
9+
- Scope targets the two non-SWE blockers: 6 structured validation-filtered rows
10+
and sparse valid/test exposure.
11+
- This task is no-training remediation/planning evidence only and does not
12+
authorize task310 training, benchmark eval, export, endpoint, or promotion.
13+
14+
## Session 1 - 2026-06-04 UTC - Accepted by worker_4
15+
16+
- Created branch
17+
`intern_nemotron_worker_4/task332_qwen_all_sft_structured_split_policy_remediation_s1`
18+
from required `origin/main`
19+
`410c2247fc5e09e6ad831bdee1628830b97fbd89`.
20+
- Imported task332 docs from lead branch commit
21+
`bbbf19df7ea7dad3fc644588f1e84240c464febe`.
22+
- Scope remains no-training remediation evidence only: exact structured
23+
filtered-row disposition, deterministic split policy or fail-closed blocker,
24+
task-owned checksums, decontam/no-AIME proof, and dependency on task331.
25+
- Boundaries: no training, eval, export, endpoint, promotion, task255 reuse,
26+
shared deletion, merge, self-merge, direct main push, or mutation of task329
27+
artifacts.

‎workspace/tasks/task332_qwen_all_sft_structured_split_policy_remediation_s1/task_knowledge.md‎

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,15 @@
1+
# task332_qwen_all_sft_structured_split_policy_remediation_s1 - task knowledge
2+
3+
<!-- METADATA:SESSION=83 -->
4+
5+
1. task329/#392 merged evidence shows valid/test splits are agentic-only; this
6+
is not accepted as an expanded all-SFT training contract.
7+
2. task329 packing receipts show 6 `instruction-following-structured`
8+
validation-filtered rows; they require exact row-level disposition.
9+
3. A positive task332 result still depends on task331 if SWE remains at zero
10+
supervised tokens.
11+
4. Any later combined packed contract must be separately lead-gated before
12+
task310 training or benchmark eval.
13+
5. Worker_4 branch for this task is
14+
`intern_nemotron_worker_4/task332_qwen_all_sft_structured_split_policy_remediation_s1`
15+
from `origin/main` `410c2247fc5e09e6ad831bdee1628830b97fbd89`.

0 commit comments

Comments
 (0)