Condense Qwen3-8B trace documentation into bullets

nl2bash_v2/README.md

@@ -129,4 +129,7 @@ The task Parquet preserves pinned references to the existing v1 image bundle; no
 
 ## Additional Qwen3-8B thinking rollouts
 
-[Separate 400-task contribution](traces/qwen3_8b_thinking_20261021.md):1,600 full trajectories with raw verifier rewards, four attempts per retainedV2 task; 900 successful attempts and 300 tasks solved. This is a subset selected from tasks with historical Coder success. Its counts are separate from all Coder-only tables and selection evidence above.
+- **400 retained V2 tasks → 4 attempts/task → 1,600 full trajectories.**
+- **900 verifier-positive attempts; 300/400 tasks solved (75% observed pass@4).**
+- Subset of tasks previously solved by Coder; counts are separate from the Coder statistics above.
+- [Trace files, settings and caveats](traces/qwen3_8b_thinking_20261021.md).

nl2bash_v2/traces/qwen3_8b_thinking_20261021.md

@@ -1,52 +1,52 @@
-# Qwen3-8B thinking: 400-task NL2Bash V2 contribution
+# Qwen3-8B thinking: NL2Bash V2 traces
 
-Adds **1,600 full attempts on 400 existing retained V2 tasks**, with four attempts per task. All attempts are included, including failures. Task content hashes match the existing Coder trace metadata at `7146b8548f19c3474827162969b270ddcae76b5f`.
+**400 retained V2 tasks → Qwen3-8B thinking × 4 attempts → 1,600 ATIF trajectories**
 
-| Coverage | Result |
-|---|---:|
-| Verifier-positive attempts |900/1,600 (56.25%)|
-| Verifier-positive attempts without a recorded exception |898|
-| Tasks solved in four attempts |300/400 (75%)|
-| Tasks with zero successful attempts |100|
-| Agent timeouts, raw grades preserved |48 (46 reward 0 ; 2 reward 1)|
+## Results
 
-**Selection matters:** these 400 tasks come from the 938 tasks already solved by the historical 30B teacher. Historical 30B observed pass@4 on this same subset is 100% by construction. The 75% figure is not an estimate on the entire original NL2Bash pool or a controlled 8B-vs-30B comparison.
+- **900/1,600 verifier-positive attempts (56.25%).**
+- **300/400 tasks solved (75% observed pass@4); 100 unsolved.**
+- **898 positive attempts without a recorded exception:** filter `reward == 1` and parsed `result.exception_info == null`.
+- **48 agent timeouts:** 46 reward 0, 2 reward 1; raw grades preserved.
 
-## Files and schema
+## Files and contents
 
-- [Parquet](qwen3_8b_thinking_20261021.parquet): the same 14 column names and Arrow types as `coder.parquet`.
-- [Manifest](qwen3_8b_thinking_20261021.manifest.json): exact revisions, settings, file SHA256, selection and per-task hashes/counts.
+- [Parquet](qwen3_8b_thinking_20261021.parquet): one row per attempt; same 14 columns and Arrow types as `coder.parquet`.
+- [Manifest](qwen3_8b_thinking_20261021.manifest.json): pinned revisions, full settings, checksums, selection and per-task coverage.
+- `trajectory`: original **ATIF-v1.8 JSON string**, including messages, reasoning, tool calls and observations.
+- `result`: trial JSON with verifier results and exceptions; operational paths, endpoints and credential-valued metadata sanitized.
+- `reward` / `verifier_reward`: original binary verifier grade; no imputation.
+- `attempt_id` / `trial`: native trial UUID / name. `source_split` / `source_shard`: this collection / batch.
+- `trajectory_found` and `retained_v2`: true for every row. Task hashes match the published Coder metadata.
 
-`trajectory` is the original native **ATIF-v1.8 JSON string** with messages, reasoning, tool calls and observations. It is not a pass-count summary. `result` is the companion trial JSON containing the verifier result and exception information. `reward` and `verifier_reward` both equal the original raw binary verifier reward; no missing reward was imputed. `trajectory_found` and `retained_v2` are true for every row.
+## Reading the results
 
-All 1,600 histories pass Harbor's native ATIF validator. All 900 successful histories contain reasoning, structured tool calls and observations. Among the 700 failed attempts, 26 have no structured tool call; these are retained rather than fabricated or discarded. Timeout histories end at their enforced cutoff.
+- **Selection bias:** these 400 tasks come from 938 previously solved by the 30B teacher. Historical 30B observed pass@4 here is 100% by construction; 75% describes this subset, not the original NL2Bash pool or a controlled model comparison.
+- **All attempts retained:** all 1,600 histories pass Harbor's ATIF validator. All 900 positive histories contain reasoning, structured tool calls and observations; 26 failed attempts have no structured tool call.
+- **Timeouts end at the cutoff:** a positive verifier grade does not guarantee normal termination. The 898-attempt filter above excludes recorded exceptions.
 
-Two timed-out attempts have raw verifier reward 1; their grades are preserved, not rewritten to 0. For a training view requiring both a positive verifier grade and no recorded exception, select reward 1 rows whose parsed `result.exception_info` is null: **898 attempts**. A positive verifier grade alone is not a claim of normal termination.
+## Run settings
 
-`attempt_id` is the original trial UUID; `trial` is its native trial name. `source_split` identifies this contribution collection, while `source_shard` identifies its native batch. These fields do not claim historical Coder-run provenance. Operational host paths, deployment endpoints and any credential-valued metadata are sanitized in `result`; trajectory text and verifier evidence are preserved. Raw originals remain with the contributor.
+- `Qwen/Qwen3-8B`, thinking enabled, BF16; exact revision in the manifest.
+- Harbor 0.24.0 → Terminus-2 2.0.0 → **Docker** with published fixtures/setup; historical runs used Apptainer. Task/verifier payloads unchanged.
+- Context 32,768; per-call output cap 8,192; temperature 0.6; top_p 0.95; top_k 20; min_p 0.
+- One A100-80GB vLLM server; 16 active agents / 20 trial slots; four independent attempts/task; no automatic trial retries.
 
-## Evaluation settings
-
-- Model: `Qwen/Qwen3-8B@b968826d9c46dd6066d109eabc6255188de91218`, thinking enabled, BF16.
-- Harbor 0.24.0 / Terminus-2 2.0.0, **Docker** rather than the historical Apptainer harness. Published fixtures are staged and their setup script executed before the agent; task/verifier payloads remain unchanged.
-- Context 32,768; explicit per-call output cap 8,192; temperature 0.6, top_p 0.95, top_k 20, min_p 0.
-- 16 active agents / 20 trial slots on one A100-80GB vLLM server; four independent attempts per task; no automatic trial retries.
-- Dataset task source: `1ca849c2128cf9ba9090a11514378c3a8277977a`; per-task hashes verified against this PR's public base. This contribution is scoped to V2 and does not alter task selection, V3 work, or Coder summaries.
-
-## Load and check
+## Load
 
 ```python
 import json
 import pyarrow.parquet as pq
 from huggingface_hub import hf_hub_download
 
-p = hf_hub_download("FWeindel/validated-tasks",
-    "nl2bash_v2/traces/qwen3_8b_thinking_20261021.parquet", repo_type="dataset")
+p = hf_hub_download(
+    "FWeindel/validated-tasks",
+    "nl2bash_v2/traces/qwen3_8b_thinking_20261021.parquet",
+    repo_type="dataset", revision="refs/pr/30",  # Omit revision after merge.
+)
 t = pq.read_table(p)
 assert t.num_rows == 1600
 assert sum(t.column("reward").to_pylist()) == 900
 trajectory = json.loads(t.column("trajectory")[0].as_py())
 result = json.loads(t.column("result")[0].as_py())
 ```
-
-While this PR is open, pass `revision="refs/pr/<number>"` to download from the PR instead of main.