Merge SETA recovery into task selection

seta_v2/README.md

@@ -2,7 +2,7 @@
 
 ## Dataset stats
 
-- **Tasks:** 4,567 native tasks; 1,414 omitted upstream, 46 excluded in v1. **3,107 v1 tasks → 2,150 kept**, all previously patched; **957 temporarily excluded** in v2.
+- **Tasks:** 4,567 native tasks; 1,414 omitted upstream, 46 excluded in v1. **3,107 v1 tasks → 2,205 kept**, including 55 newly retained after pass@16 recovery; **902 temporarily excluded**.
 - **Images:** reuse **10 of v1's 11 cached Apptainer environments**. No additional images or task repairs in v2; original unique image count was not recorded.
 
 ## Dataset characteristics
@@ -20,30 +20,28 @@
 
 Intended benchmark transfer: **Terminal-Bench 2.0, OpenThoughts-TBLite 2.0, DS-1000 mini-200 and BixBench CLI corrections @ Pi**, through terminal workflows, Python data analysis and executable verification.
 
-## Partial-result selection
+## Merged partial-result selection
 
-Snapshot: **15 October 2026, 20:57 UTC**. Keep a task as soon as either model records a successful attempt, even if other attempts are unfinished. **811 kept tasks have incomplete results; 1,339 have all 16 graded attempts per model.**
+The published task list combines the frozen 15 October 2026, 20:57 UTC snapshot with the completed `2026-10-15-pass16-recovery` run. A task is kept when either model has at least one verifier reward of 1 across those results. The original snapshot remains recorded in `metadata/selection.json`; recovery results are stored separately so overlapping attempt counts are not added or double-counted.
 
 | Model | Tasks solved at least once |
 | --- | ---: |
-| Qwen3-30B-A3B-Instruct-2507 | 1,991 |
-| Qwen3-30B-A3B-Thinking | 1,356 |
-| Either model | **2,150** |
+| Qwen3-30B-A3B-Instruct-2507 | 2,137 |
+| Qwen3-30B-A3B-Thinking | 1,459 |
+| Either model | **2,205** |
 
-| Temporary exclusion reason | Tasks |
+| Still temporarily excluded | Tasks |
 | --- | ---: |
+| Incomplete results, no success recorded | 509 |
 | Complete 16 attempts per model, no success | 364 |
-| Incomplete results, no success recorded yet | 564 |
-| No scored results yet | 29 |
-| Total | **957** |
+| No scored results | 29 |
+| Total | **902** |
 
-Missing results remain unknown, not failures. The 593 exclusions with missing/incomplete results can be reconsidered as runs finish. A lack of success does not establish a task defect.
+The recovery newly retained 55 tasks: 50 first successes for Instruct and 5 for Thinking. Another 194 existing tasks gained a first observed success for one model; model totals count each task once. Missing verifier rewards remain unknown, not failures.
 
-Build, reference, no-op and rule-based checks carry over from [v1](../seta_v1/README.md). Evaluated hashes match v1; kept payloads are unchanged.
+## Successful attempts in the original snapshot
 
-## Successful attempts among kept tasks
-
-These are observed success counts from up to 16 attempts per model, not complete pass@16 estimates. Zero for one model can include missing results; the other model solved the task.
+The table below records the frozen 15 October snapshot only (2,150 kept tasks). It is retained for historical comparison and is not a distribution for the merged 2,205-task selection. Recovery attempts and raw rewards are available in `runs/2026-10-15-pass16-recovery/`; `tasks.parquet` has separate `recovery_*` columns for each task.
 
 | Successful attempts | Instruct tasks | Thinking tasks |
 | --- | ---: | ---: |
@@ -67,13 +65,15 @@ These are observed success counts from up to 16 attempts per model, not complete
 
 ## Files and selecting tasks
 
-`tasks.parquet` contains kept payloads and each model's successes, graded attempts and missing attempts. `metadata/selection.json` records every v1 task's results, selection reason and the pinned v1 revision. Excluded payloads remain in v1; no duplicate archive Parquet is uploaded. `patch.py` reproduces this snapshot.
+`tasks.parquet` contains the 2,205 kept payloads, frozen-snapshot metrics, and separate per-task recovery attempt, success and missing-reward counts. `metadata/selection.json` records every v1 task's results, selection reason and the pinned v1 revision. Excluded payloads remain in v1; no duplicate archive Parquet is uploaded. `patch.py` reproduces the merged task list from the pinned source tasks and `metadata/selection.json`.
 
 ```python
 import pyarrow.parquet as pq
-# Observed exactly one Instruct success, including incomplete runs.
-subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])
-# Add ("instruct_graded_attempts", "=", 16) to require a complete Instruct run.
+# At least one Instruct success in either the frozen snapshot or recovery tranche.
+t = pq.read_table("tasks.parquet")
+subset = t.filter(
+    (t["instruct_successes"] + t["recovery_instruct_successes"]) > 0
+)
 pq.write_table(subset, "selected.parquet", compression="zstd")
 ```
 

seta_v2/patch.py

@@ -1,4 +1,4 @@
-"""Reproduce the frozen SETA v2 partial-result selection. Requires pyarrow and huggingface_hub."""
+"""Reproduce the merged SETA v2 partial-result selection. Requires pyarrow and huggingface_hub."""
 import argparse,copy,hashlib,json
 from pathlib import Path
 import pyarrow as pa
@@ -14,6 +14,9 @@ for prefix in ['instruct','thinking']:
  for suffix in ['successes','graded_attempts','missing_attempts']:
   schema=schema.append(pa.field(prefix+'_'+suffix,pa.int64()))
 schema=schema.append(pa.field('target_attempts_per_model',pa.int64()))
+for prefix in ['recovery_instruct','recovery_thinking']:
+ for suffix in ['attempts','successes','missing_rewards']:
+  schema=schema.append(pa.field(prefix+'_'+suffix,pa.int64()))
 for r in t.to_pylist():
  s=m['tasks'][r['path']];assert r['content_sha256']==s['content_sha256']
  if not s['kept']:continue
@@ -22,9 +25,16 @@ for r in t.to_pylist():
   assert v['graded_attempts']==sum(x is not None for x in v['rewards'])
   assert v['missing_attempts']==16-v['graded_attempts']
   for suffix,key in [('successes','successful_attempts'),('graded_attempts','graded_attempts'),('missing_attempts','missing_attempts')]:r[prefix+'_'+suffix]=v[key]
- assert r['instruct_successes'] or r['thinking_successes']
+ recovery=s.get('recovery',{})
+ for prefix in ['instruct','thinking']:
+  v=recovery.get(prefix,{})
+  r['recovery_'+prefix+'_attempts']=v.get('attempts',0)
+  r['recovery_'+prefix+'_successes']=v.get('successful_attempts',0)
+  r['recovery_'+prefix+'_missing_rewards']=v.get('missing_verifier_rewards',0)
+ assert (r['instruct_successes'] or r['thinking_successes'] or
+         r['recovery_instruct_successes'] or r['recovery_thinking_successes'])
  r['target_attempts_per_model']=16;rows.append(r)
-assert len(rows)==2150
+assert len(rows)==2205
 names={r['path'] for r in rows};images=json.loads(t.schema.metadata[b'ot.images.v1']);images['tasks']={k:v for k,v in images['tasks'].items() if k in names};assert set(images['tasks'])==names
 bundles={e['bundle'] for v in images['tasks'].values() for e in v['environments']};images['bundles']={k:v for k,v in images['bundles'].items() if k in bundles};images['missing']=[]
 meta={b'ot.images.v1':json.dumps(images).encode(),b'ot.release.v2':json.dumps({k:v for k,v in m.items() if k!='tasks'}).encode()}

seta_v2/runs/2026-10-15-pass16-recovery/README.md

@@ -4,4 +4,4 @@ This folder contains the completed remaining-attempt recovery tranche for Qwen3-
 
 The missing-reward audit contains 764 final trial attempts with their original exception messages: 524 harness failures, 63 supported agent failures, 32 verifier timeouts and 145 unresolved errors. The agent failures include terminal exits and terminal kills.
 
-This is the recovery tranche, not an independent full pass@16 estimate. The existing SETA v2 selection remains the frozen 2026-10-15 20:57 UTC snapshot. See `run.json` for report hashes, job IDs, counts and provenance.
+This is the recovery tranche, not an independent full pass@16 estimate. The original SETA v2 selection remains available as the frozen 2026-10-15 20:57 UTC baseline in `metadata/selection.json`. The published task list now merges this recovery evidence under the existing rule and retains 55 additional tasks. Per-task recovery counts are separate from baseline counts to prevent double-counting overlapping attempts. See `run.json` for report hashes, job IDs, counts and provenance.

seta_v2/runs/2026-10-15-pass16-recovery/run.json

@@ -913,5 +913,17 @@
       "bytes": 2356,
       "sha256": "f91ae6b6d872e76c1724d8bea2d410f9c932c9c145a86e6cc4f1ef441e0758e1"
     }
-  ]
+  ],
+  "selection_merge": {
+    "as_of_utc": "2026-10-16T10:10:10+00:00",
+    "rule": "Keep task if either model has verifier reward 1 in frozen snapshot or this recovery run.",
+    "newly_retained_tasks": 55,
+    "published_kept_tasks": 2205,
+    "published_excluded_tasks": 902,
+    "solved_at_least_once": {
+      "qwen3-30b-instruct-2507": 2137,
+      "qwen3-30b-thinking": 1459
+    },
+    "baseline_attempt_counts_preserved": true
+  }
 }

seta_v2/tasks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1d447ca273c69e00b92abb677e6ae710d8e13bf345808b96f44c849ace5aa505
-size 224982161
+oid sha256:927df4ad0790782f500575a560a792e13a2ca3a31b52ff846408e0709c040653
+size 249061328