Merge PR #321: task242 Qwen3-4B V10 planner smoke

Self-merge approved by lead after #320 merged. Scope: Qwen3-4B V10 planner smoke wiring, fail-closed decontam checks, task-owned /root sync guard, and 30B hold. No training/live eval/sync/30B launched.

Author: songCNMS

src/nemotron/recipes/super3/milestones/m1_agentic_sft/plan_qwen_scaleup_run.py

@@ -609,6 +609,188 @@ def test_scaleup_planner_can_emit_hard_math_recurrence_v9_data_prep(tmp_path) ->
     assert "pack_size=8192" in local_script
 
 
+def test_scaleup_planner_can_emit_qwen4b_v10_pilot_bundle(tmp_path) -> None:
+    corpus_path = tmp_path / "aime25_hmmt_math_corpus.jsonl"
+    corpus_path.write_text(
+        '{"id":"heldout-smoke","prompt":"Held-out AIME/HMMT/MATH prompt text."}\n',
+        encoding="utf-8",
+    )
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "task242_qwen4b_v10"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--run-name",
+            "task242_qwen4b_v10_pilot",
+            "--qwen4b-v10-pilot",
+            "--math-sidecar-m0-input-dir",
+            "/data/full_m0",
+            "--math-decontaminate-against-corpus",
+            str(corpus_path),
+            "--pack-size",
+            "8192",
+            "--seq-length",
+            "8192",
+        ]
+    )
+    manifest = build_manifest(args)
+    local_script = render_local_data_prep_script(manifest)
+    sync_script = render_sync_script(manifest)
+    remote_script = render_remote_train_script(manifest)
+    eval_script = render_eval_script(manifest)
+    report = render_report(manifest)
+
+    assert manifest["pilot_profile"] == "qwen3_4b_v10_aime"
+    assert manifest["task"] == "task242_qwen_aime_v10_planner_smoke_s1"
+    assert (
+        manifest["training"]["qwen_hf_model"]
+        == "/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507"
+    )
+    assert manifest["training"]["pretrained_checkpoint"] == manifest["training"]["qwen_hf_model"]
+    assert manifest["training"]["train_entrypoint"].endswith("qwen_local_train.py")
+    assert manifest["paths"]["remote_root"] == "/root/task242_qwen_aime_v10_planner_smoke_s1"
+    assert manifest["data"]["math_supervision_strategy"] == "hard_math_runlength_dp_v10"
+    assert manifest["data"]["math_decontaminate_against_corpus"] == str(corpus_path)
+    assert manifest["aime_gate"]["enabled"] is True
+    assert manifest["aime_gate"]["base_model_path"] == manifest["training"]["qwen_hf_model"]
+    assert (
+        manifest["aime_gate"]["candidate_ft_output_path"]
+        == f"{manifest['paths']['remote_run_root']}/checkpoints"
+    )
+    assert (
+        manifest["aime_gate"]["non_regression_rule"]["pass_condition"]
+        == "ft_exact_normalized_accuracy >= base_exact_normalized_accuracy"
+    )
+    assert "held until Qwen3-4B V10 FT AIME25 smoke" in manifest["aime_gate"]["scale_hold"][
+        "qwen30b_8gpu"
+    ]
+
+    assert "--math-supervision-strategy hard_math_runlength_dp_v10" in local_script
+    assert "--math-v10-hard-verified-full-solution-weight 1.0" in local_script
+    assert "requires a non-empty decontamination corpus" in local_script
+    assert "TASK242_DECONTAM_CORPUS_PLACEHOLDER" in local_script
+    assert f"--decontaminate-math-against-corpus {corpus_path}" in local_script
+    assert "--math-sidecar-m0-input-dir /data/full_m0" in local_script
+    assert "V10 pilot sync must target /root" in sync_script
+    assert "does not delete /mnt/cephfs/data/processing/lei.song" in sync_script
+    assert "Qwen3-4B-Instruct-2507" in remote_script
+    assert "task242_task242_qwen4b_v10_pilot" in remote_script
+    assert "/root/task242_qwen_aime_v10_planner_smoke_s1" in remote_script
+    assert "V10 AIME gate contract" in eval_script
+    assert "ft_exact_normalized_accuracy >= base_exact_normalized_accuracy" in eval_script
+    assert "Qwen3-4B V10 AIME Gate" in report
+    assert "30B/8-GPU hold" in report
+
+
+def test_scaleup_planner_v10_fails_closed_without_decontamination_corpus(tmp_path) -> None:
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "scaleup"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--qwen4b-v10-pilot",
+        ]
+    )
+
+    with pytest.raises(ValueError, match="hard_math_runlength_dp_v10 requires"):
+        build_manifest(args)
+
+
+def test_scaleup_planner_v10_rejects_missing_decontamination_corpus_path(tmp_path) -> None:
+    missing_corpus = tmp_path / "missing_corpus.jsonl"
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "scaleup"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--qwen4b-v10-pilot",
+            "--math-decontaminate-against-corpus",
+            str(missing_corpus),
+        ]
+    )
+
+    with pytest.raises(ValueError, match="decontamination corpus is missing"):
+        build_manifest(args)
+
+
+def test_scaleup_planner_v10_rejects_empty_decontamination_corpus_path(tmp_path) -> None:
+    empty_corpus = tmp_path / "empty_corpus.jsonl"
+    empty_corpus.write_text("", encoding="utf-8")
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "scaleup"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--qwen4b-v10-pilot",
+            "--math-decontaminate-against-corpus",
+            str(empty_corpus),
+        ]
+    )
+
+    with pytest.raises(ValueError, match="decontamination corpus is empty"):
+        build_manifest(args)
+
+
+def test_scaleup_planner_v10_holds_30b_until_pilot_gate(tmp_path) -> None:
+    corpus_path = tmp_path / "aime25_hmmt_math_corpus.jsonl"
+    corpus_path.write_text(
+        '{"id":"heldout-smoke","prompt":"Held-out AIME/HMMT/MATH prompt text."}\n',
+        encoding="utf-8",
+    )
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "scaleup"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--qwen-hf-model",
+            "/models/Qwen3-30B-A3B-Instruct-2507",
+            "--pretrained-checkpoint",
+            "/checkpoints/qwen3-30b-a3b-bridge",
+            "--math-supervision-strategy",
+            "hard_math_runlength_dp_v10",
+            "--math-decontaminate-against-corpus",
+            str(corpus_path),
+        ]
+    )
+
+    with pytest.raises(ValueError, match="30B planning is held"):
+        build_manifest(args)
+
+
+def test_scaleup_planner_v10_preserves_30b_entrypoint_after_gate_override(tmp_path) -> None:
+    corpus_path = tmp_path / "aime25_hmmt_math_corpus.jsonl"
+    corpus_path.write_text(
+        '{"id":"heldout-smoke","prompt":"Held-out AIME/HMMT/MATH prompt text."}\n',
+        encoding="utf-8",
+    )
+    args = build_parser().parse_args(
+        [
+            "--output-dir",
+            str(tmp_path / "scaleup"),
+            "--repo-dir",
+            str(tmp_path / "repo"),
+            "--qwen-hf-model",
+            "/models/Qwen3-30B-A3B-Instruct-2507",
+            "--pretrained-checkpoint",
+            "/checkpoints/qwen3-30b-a3b-bridge",
+            "--math-supervision-strategy",
+            "hard_math_runlength_dp_v10",
+            "--math-decontaminate-against-corpus",
+            str(corpus_path),
+            "--allow-v10-30b-scale",
+        ]
+    )
+    manifest = build_manifest(args)
+
+    assert manifest["training"]["train_entrypoint"] == QWEN30B_A3B_TRAIN_ENTRYPOINT
+    assert manifest["aime_gate"]["scale_hold"]["allow_v10_30b_scale"] is True
+
+
 def test_scaleup_planner_plumbs_math_decontamination_flags_through_local_script(tmp_path) -> None:
     """Regression: prepare_m1_agentic_sft.py requires
     --decontaminate-math-against-corpus for V7+ strategies (or

tests/recipes/super3/test_m1_agentic_qwen_scaleup_plan.py

@@ -1,13 +1,14 @@
 # intern_nemotron_worker_2 - status
 
-<!-- METADATA:STATUS=Working,TASK=task217_mamba_causal_conv_train_stack_unblock_probe_s1,ROLE=worker,TEAM_ID=nemotron -->
+<!-- METADATA:STATUS=Working,TASK=task242_qwen_aime_v10_planner_smoke_s1,ROLE=worker,TEAM_ID=nemotron -->
 
 | Field | Value |
 |------|-----|
 | Name | intern_nemotron_worker_2 |
 | Status | Working |
 | Role | worker |
 | Team | nemotron |
-| Current Task | task217_mamba_causal_conv_train_stack_unblock_probe_s1 |
-| PR | https://github.com/songCNMS/Nemotron/pull/316 |
+| Current Task | task242_qwen_aime_v10_planner_smoke_s1 |
+| PR | https://github.com/songCNMS/Nemotron/pull/321 |
 | Session | 2 |
+| Last Update | 2026-06-01T16:03:24Z: Opened PR #321 for V10 planner/Qwen3-4B smoke bundle; waiting for review/gate, no training or 30B scale launched. |

workspace/interns/intern_nemotron_worker_2/status.md

@@ -0,0 +1,48 @@
+# task242_qwen_aime_v10_planner_smoke_s1 - Qwen AIME V10 planner and smoke scripts
+
+<!-- METADATA:STATUS=ReadyForPR,ASSIGNEE=intern_nemotron_worker_2,SESSION=2 -->
+
+## Background
+
+The team must use Qwen3-4B for cheap pilot/debug before spending 30B/8-GPU scale. Existing project rules say code/debug runs happen on remote node `NemTron`, code must be synced to `/root` before debug, and the small debug checkpoint is `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`.
+
+## Goal
+
+Expose the V10 data strategy through planner/training scripts and produce safe smoke launch scripts for a Qwen3-4B pilot. The planner must enforce the same base-vs-FT eval gate before any 30B scaling is proposed.
+
+## Scope
+
+- Own planner changes in `src/nemotron/recipes/super3/milestones/m1_agentic_sft/plan_qwen_scaleup_run.py` and related training-planner surfaces.
+- Add or update generated-script support for the V10 strategy created by worker_1.
+- Provide a Qwen3-4B pilot plan using `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`.
+- Preserve Qwen3-30B-A3B train-entrypoint routing for any later scale-up, but do not launch or spend 30B/8-GPU scale in this task.
+- Make smoke scripts record local CPU prep, sync-to-`/root` on NemTron, and safe shared-storage behavior.
+
+## Boundaries
+
+- Do not push `main` or self-merge.
+- Do not launch full 30B/8-GPU training.
+- Do not delete existing files in `/mnt/cephfs/data/processing/lei.song`.
+- Do not judge pilot success without worker_3 same-harness base score.
+- Coordinate with worker_1 if planner changes need data-strategy names that are still in review.
+
+## Expected Output
+
+- Worker branch: `intern_nemotron_worker_2/task242_qwen_aime_v10_planner_smoke_s1`.
+- PR to `main` after local validation.
+- A planner report in this task directory with generated command examples, pilot resource assumptions, expected artifact paths, and 30B scale-up hold conditions.
+- Mailbox report with branch, head SHA, PR URL, files touched, tests/checks run, and resource blockers.
+
+## Acceptance Criteria
+
+- Planner can generate a Qwen3-4B V10 pilot bundle without requiring 30B weights.
+- Planner manifest records base model path, candidate FT output path, corrected AIME evaluator config, and base-vs-FT non-regression rule.
+- Scripts fail closed if decontamination corpus is missing for V10 hard-math data.
+- First measurable go/no-go gate is encoded in the run plan: compare Qwen3-4B base and Qwen3-4B V10 FT under the same corrected AIME25 smoke protocol before any 30B proposal.
+
+## Assignment
+
+- Team: `nemotron`
+- Team lead: `intern_nemotron_lead`
+- Worker: `intern_nemotron_worker_2`
+- Depends on: worker_1 V10 sidecar interface, worker_3 baseline/eval protocol, worker_5 runbook verification

workspace/tasks/task242_qwen_aime_v10_planner_smoke_s1/README.md

@@ -0,0 +1,23 @@
+# task242_qwen_aime_v10_planner_smoke_s1 - History Log
+
+<!-- METADATA:SESSION=2 -->
+
+## Session 0 - Assigned
+
+- Created by `intern_nemotron_lead` for `intern_nemotron_worker_2`.
+- Initial focus: V10 planner support, Qwen3-4B pilot scripts, and explicit hold on 30B/8-GPU scale until the pilot non-regression gate is satisfied.
+
+## Session 1 - 2026-06-01 UTC - Accepted by worker
+
+- Worker `intern_nemotron_worker_2` accepted the task.
+- Created branch `intern_nemotron_worker_2/task242_qwen_aime_v10_planner_smoke_s1` from current `origin/main`.
+- Imported task docs from `origin/intern_nemotron_lead/session1-recovery-task-docs` at `116a2f3`.
+
+## Session 2 - 2026-06-01 UTC - V10 planner and smoke bundle
+
+- Added planner wiring for `hard_math_runlength_dp_v10`, including V10 weights, Qwen3-4B pilot defaults, same-harness AIME gate manifest fields, and a 30B/8-GPU hold unless `--allow-v10-30b-scale` is explicitly supplied after the 4B gate passes.
+- Added fail-closed V10 decontamination validation: missing, non-file, empty, skip-check, and unapproved 30B paths are rejected at planning time; generated local data-prep scripts also reject the task242 placeholder corpus marker before running.
+- Generated task-owned pilot bundle under `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot` with local data-prep, `/root` NemTron sync, remote train, eval dry-run scripts, and manifest/report artifacts.
+- Local checks passed: py_compile, focused planner pytest (`29 passed`), ruff, and `git diff --check`.
+- Opened PR #321 to `main`: https://github.com/songCNMS/Nemotron/pull/321.
+- Did not run training, live eval, or 30B/8-GPU scale; blockers are the real held-out decontamination corpus, task241 V10 data-prep merge, and task243 same-harness base/FT AIME evidence.

workspace/tasks/task242_qwen_aime_v10_planner_smoke_s1/history_log.md

@@ -0,0 +1,84 @@
+# task242 planner report
+
+<!-- METADATA:STATUS=ReadyForPR,SESSION=2 -->
+
+## Summary
+
+- Added planner support for `hard_math_runlength_dp_v10`.
+- Added a Qwen3-4B V10 pilot profile using `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`.
+- Encoded the same-harness AIME25 base-vs-FT non-regression gate in the manifest before any 30B/8-GPU scale is allowed.
+- Generated a task-owned smoke bundle under `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot`.
+- Did not run training, live eval, or any 30B/8-GPU job.
+
+## Generated command
+
+```bash
+PYTHONPATH=src python src/nemotron/recipes/super3/milestones/m1_agentic_sft/plan_qwen_scaleup_run.py \
+  --qwen4b-v10-pilot \
+  --run-name task242_qwen4b_v10_pilot \
+  --math-sidecar-m0-input-dir /work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/task241_v10_math_sidecar_m0_PENDING \
+  --math-decontaminate-against-corpus /work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/aime25_hmmt_math_heldout_decontam_corpus.PLACEHOLDER.jsonl \
+  --pack-size 8192 \
+  --seq-length 8192 \
+  --num-shards 8 \
+  --max-train-per-dataset 100 \
+  --max-val-per-dataset 25 \
+  --math-sidecar-max-records-per-env 500 \
+  --math-sidecar-max-val-shadow-per-env 25 \
+  --epochs 0.05 \
+  --eval-interval 20 \
+  --save-interval 20 \
+  --overwrite
+```
+
+## Generated artifacts
+
+- Manifest: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/scaleup_manifest.json`
+- Planner report: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/report.md`
+- Local data-prep script: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/run_local_data_prep.sh`
+- NemTron sync script: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/sync_to_nemtron.sh`
+- NemTron train script: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/run_nemtron_train.sh`
+- Eval dry-run script: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot/run_eval_basket_dry_run.sh`
+
+## Pilot contract
+
+- Pilot model/checkpoint/tokenizer path: `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`
+- Math strategy: `hard_math_runlength_dp_v10`
+- Default V10 weights: hard verified full-solution `1.0`, broad verified `0.0`, final-answer aux `0.0`, format repair `0.0`
+- Local output root: `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot`
+- NemTron remote root: `/root/task242_qwen_aime_v10_planner_smoke_s1`
+- Remote run root: `/root/task242_qwen_aime_v10_planner_smoke_s1/task242_qwen_aime_v10_4b_pilot`
+- Train tmux session name: `task242_task242_qwen4b_v10_pilot`
+
+## Safety checks
+
+- V10 manifest creation fails if `--math-decontaminate-against-corpus` is missing, not a file, or empty.
+- Generated local data-prep script refuses the task242 placeholder corpus marker before running M0/M1 data prep.
+- Generated sync script refuses non-`/root/*` V10 pilot remote roots.
+- Generated sync script only removes the task-owned `/root/task242_qwen_aime_v10_planner_smoke_s1/...` paths and prints that it does not delete `/mnt/cephfs/data/processing/lei.song`.
+
+## AIME gate
+
+- Gate id: `qwen3_4b_v10_aime25_same_harness_non_regression`
+- Base model path: `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`
+- Candidate FT output path: `/root/task242_qwen_aime_v10_planner_smoke_s1/task242_qwen_aime_v10_4b_pilot/checkpoints`
+- Corrected evaluator config: `src/nemotron/recipes/super3/stage3_eval/config/m1_corrected_math_comparison.yaml`
+- Pilot protocol: AIME25 held-out prompts, 1 repeat per problem, `8192` max tokens, `/v1/chat/completions`, `temperature=0.0`, `top_p=1e-5`
+- Non-regression rule: `ft_exact_normalized_accuracy >= base_exact_normalized_accuracy`
+- Required diagnostics: numerator, denominator, parsed count, finish reasons, per-problem rows
+- 30B hold: Qwen3-30B-A3B / 8-GPU planning is refused unless `--allow-v10-30b-scale` is explicitly supplied after the Qwen3-4B same-harness gate is documented as passing.
+
+## Checks
+
+- `python -m py_compile src/nemotron/recipes/super3/milestones/m1_agentic_sft/plan_qwen_scaleup_run.py` passed.
+- `PYTHONPATH=src pytest -q tests/recipes/super3/test_m1_agentic_qwen_scaleup_plan.py` passed: 29 passed.
+- `ruff check src/nemotron/recipes/super3/milestones/m1_agentic_sft/plan_qwen_scaleup_run.py tests/recipes/super3/test_m1_agentic_qwen_scaleup_plan.py` passed.
+- `git diff --check` passed.
+- Verified `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507` exists in this worker environment.
+
+## Blockers and residual risk
+
+- Real AIME25/HMMT/MATH held-out decontamination corpus is not visible in this worker workspace. The generated bundle uses a task-owned placeholder only to materialize paths; the generated local data-prep script fails closed until it is replaced.
+- The task241 V10 data-prep interface is on `origin/intern_nemotron_worker_1/task241_qwen_aime_v10_sidecar_data_s1` and is not on `origin/main` at this report time. The planner is wired to the worker_1 interface, but local data prep will not run on main until that code lands or is explicitly combined.
+- No same-harness Qwen3-4B base or FT AIME25 live score was produced here. Task243 owns that gate; this task only records the required manifest contract and script-side hold.
+- No training, sync, or live eval commands were executed.

workspace/tasks/task242_qwen_aime_v10_planner_smoke_s1/planner_report.md

@@ -0,0 +1,10 @@
+# task242_qwen_aime_v10_planner_smoke_s1 - Task Knowledge
+
+<!-- METADATA:SESSION=2 -->
+
+## Knowledge Entries
+
+1. Qwen3-4B pilot/debug path is `/mnt/cephfs/data/stable/models/Qwen/Qwen3-4B-Instruct-2507`.
+2. Any later Qwen3-30B-A3B scale-up must remain held until Qwen3-4B pilot AIME25 is non-regressing or identifies a concrete fix.
+3. Project rule: code/debug runs happen on `NemTron`, code syncs to `/root` before debug, and shared `/mnt/cephfs/data/processing/lei.song` files must not be deleted.
+4. Task242 generated smoke bundle path is `/work-agents/intern_nemotron_worker_2/outputs/task242_qwen_aime_v10_4b_pilot`; it is intentionally not a completed data/training artifact because the real held-out decontamination corpus and task241 V10 data-prep merge are still required.