Endless Terminals v2: retain 1,958 tasks using both teachers

endless_v2/README.md

@@ -2,9 +2,9 @@
 
 ## Dataset statistics
 
-- **2,492 original tasks → 2,144 repaired v1 tasks → 1,323 kept in v2.** All kept tasks were patched in v1; v2 changes no task payloads.
-- **821 v1 tasks excluded for now:** 818 had no successful attempt; three have no recorded results. Excluded payloads remain in [v1](../endless_v1).
-- **2,492 original build contexts → 1,834 cached Apptainer environments in v1.** V2 reuses **1,159** of those environments; no duplicate image upload.
+- **2,492 original tasks → 2,144 repaired v1 tasks → 1,958 kept in v2.** All kept tasks were patched in v1; v2 changes no task payloads. This update restores **635 tasks** using the additional Coder results.
+- **186 v1 tasks excluded:** 183 had zero successful attempts in both models; three have no recorded attempts. Their payloads and reasons are in [archive.parquet](archive.parquet). The **348 original exclusions inherited from v1 remain excluded**.
+- **2,492 original build contexts → 1,834 cached Apptainer environments in v1.** V2 reuses **1,676** of those environments; no duplicate image upload.
 
 ## Dataset characteristics
 
@@ -23,18 +23,37 @@ Intended benchmark transfer: **Terminal-Bench 2.0 and OpenThoughts-TBLite 2.0**,
 
 ## Validation and filtering
 
-- **Environment reliability and rule-based task quality:** keeps all [v1 repairs](../endless_v1), including build/setup, reference-solution and no-op verification, and static-check repairs.
-- **Pass@4 filtering:** retain tasks solved at least once by **Qwen3-30B-A3B-Instruct-2507**, using Terminus 2. Evaluated content hashes match every kept task.
-- **15 October 2026 results:** 2,141 tasks completed four attempts each; none are partial; three have no recorded attempts. Terminal execution failures count as unsuccessful attempts. Five attempts have no raw verifier reward and receive a filtering score of zero.
+- **Environment reliability and rule-based task quality:** keeps all [v1 repairs](../endless_v1), including build/setup, reference-solution and no-op verification, and static-check repairs. Inherited validation limitations remain: the conversion review had no Docker reference run, and one HTTP task has an unresolved possible shared-port collision.
+- **Either-model success filtering:** retain tasks solved at least once by **Qwen3-30B-A3B-Instruct-2507** in four attempts or **Qwen3-Coder-30B-A3B-Instruct** in eight attempts, using Terminus 2. Evaluated content hashes match every kept task. This retains **34 tasks solved by Instruct but not Coder**.
+- **15–16 October 2026 results:** each model evaluated the same **2,141 tasks**. Instruct completed **8,564 attempts**; Coder completed **17,128 attempts** in two four-attempt collections. No measured tasks have partial attempt coverage. Three v1 tasks remain unmeasured and excluded.
+- Recorded terminal execution failures count as unsuccessful attempts. **15 attempts** have no raw verifier reward—five Instruct and ten Coder—and receive a filtering score of zero. Their original verifier/error records are preserved.
 
-### Successful attempts among kept tasks
+### Teacher results: pass@k
 
-| Successful attempts | Tasks |
-| --- | ---: |
-| 1/4 | 288 |
-| 2/4 | 254 |
-| 3/4 | 261 |
-| 4/4 | 520 |
+Rates cover **all 2,141 measured tasks before teacher-based filtering**, including tasks excluded from v2. Each model is evaluated separately using its own attempt budget.
+
+| Model | Attempts per task | Correct attempts | Tasks solved at least once | Pass@1 | Pass@4 | Pass@8 |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Qwen3-30B-A3B-Instruct-2507 | 4 | 3,659 / 8,564 | 1,323 / 2,141 | 42.73% | 61.79% | — |
+| Qwen3-Coder-30B-A3B-Instruct | 8 | 11,371 / 17,128 | 1,924 / 2,141 | 66.39% | 85.43% | 89.86% |
+
+Pass@k uses the unbiased estimator from each task's recorded success and attempt counts. Coder's pass@4 above uses all eight attempts; its first four-attempt collection had **86.27% pass@4**. Instruct has no pass@8 measurement. Per-model rates and evaluation scope are saved in [metadata/pass_at_k.json](metadata/pass_at_k.json); [pass_counts.parquet](pass_counts.parquet) contains one row per measured task and model.
+
+### Successful attempts among the 1,958 kept tasks
+
+| Successful attempts | Instruct: four attempts | Coder: eight attempts |
+| --- | ---: | ---: |
+| 0 | 635 | 34 |
+| 1 | 288 | 121 |
+| 2 | 254 | 109 |
+| 3 | 261 | 127 |
+| 4 | 520 | 138 |
+| 5 | — | 171 |
+| 6 | — | 239 |
+| 7 | — | 342 |
+| 8 | — | 677 |
+
+A zero in one model's column is compatible with retention because the other model succeeded.
 
 ### Select by successful attempts
 
@@ -44,14 +63,18 @@ import pyarrow.parquet as pq
 
 path = hf_hub_download("FWeindel/validated-tasks", "endless_v2/tasks.parquet", repo_type="dataset")
 tasks = pq.read_table(path, filters=[("instruct_successes", ">=", 1), ("instruct_successes", "<=", 2)])
+# Coder counts are out of eight attempts:
+coder_tasks = pq.read_table(path, filters=[("coder_successes", ">=", 1), ("coder_successes", "<=", 2)])
 ```
 
-`metadata/selection.json` records all v1 task decisions and counts. `patch.py` reproduces the task Parquet from pinned v1 data. `instruct_graded_attempts` counts attempts assigned a filtering score, including terminal failures.
+[metadata/selection.json](metadata/selection.json) records decisions and per-model counts for all 2,144 v1 tasks. [patch.py](patch.py) reproduces the task, archive and pass-count Parquets from pinned v1 data. The `*_graded_attempts` columns count attempts assigned a filtering score, including terminal failures. `instruct_target_attempts` is four and `coder_target_attempts` is eight. The legacy `target_attempts_per_model` column remains four for compatibility with the original Instruct-only schema; it is not Coder's attempt budget.
 
 ## Traces
 
-[Compressed traces](traces/attempts.parquet) include **all 8,564 attempts**, successful and failed, across kept and excluded tasks. No conversations are missing. Each row contains the task ID, model, attempt ID, filtering reward, raw verifier reward, retention flag and full trajectory and result JSON.
+[Compressed traces](traces/attempts.parquet) include **all 25,692 attempt records**: 8,564 Instruct and 17,128 Coder, successful and failed, across kept and excluded measured tasks. Each row contains the task ID, model, attempt ID, filtering reward, raw verifier reward, current retention flag, result JSON and saved trajectory JSON when available.
+
+**One Coder attempt has no saved conversation**; its result/error record remains included and `trajectory_found` is false. All Instruct conversations are present. Missing raw verifier rewards remain null rather than being overwritten with the filtering score. [traces/manifest.json](traces/manifest.json) records counts, per-model results, missing-data counts and the trace Parquet checksum.
 
 ## Resources
 
-Historical teacher runs used **one H200, 32 CPUs per job and concurrency 10**. Fifteen jobs each took approximately **55 to 79 minutes**. Environment and image versions are inherited from [v1](../endless_v1).
+Teacher runs used **one H200, 32 CPUs per job and concurrency 10**. Each collection used fifteen jobs, with elapsed allocation times of approximately **55 to 79 minutes**. Coder's second collection started after its first completed successfully. Environment and image versions are inherited from [v1](../endless_v1).

endless_v2/archive.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab1f6de887c5670c84061e31c3ea3c0fce1ac6219490afad23fe4567df9d41a5
+size 3590866

endless_v2/metadata/pass_at_k.json

@@ -0,0 +1,48 @@
+{
+  "scope": "All 2,141 measured v1 tasks before teacher-based selection; each model uses its own attempt budget.",
+  "models": {
+    "instruct": {
+      "model": "Qwen/Qwen3-30B-A3B-Instruct-2507",
+      "tasks": 2141,
+      "attempts_per_task": 4,
+      "attempts": 8564,
+      "successful_attempts": 3659,
+      "solved_tasks": 1323,
+      "zero_success_tasks": 818,
+      "pass_rates": {
+        "pass@1": 0.42725361980383,
+        "pass@2": 0.5309045617312782,
+        "pass@3": 0.5843063988790285,
+        "pass@4": 0.6179355441382531
+      }
+    },
+    "coder": {
+      "model": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+      "tasks": 2141,
+      "attempts_per_task": 8,
+      "attempts": 17128,
+      "successful_attempts": 11371,
+      "solved_tasks": 1924,
+      "zero_success_tasks": 217,
+      "pass_rates": {
+        "pass@1": 0.6638836992059786,
+        "pass@2": 0.7814439180623207,
+        "pass@3": 0.8285180489757791,
+        "pass@4": 0.8543204110228865,
+        "pass@8": 0.8986454927603923
+      }
+    }
+  },
+  "selection_summary": {
+    "tasks": 2144,
+    "solved": 1958,
+    "complete": 2141,
+    "incomplete": 0,
+    "zero_success_both_models": 183,
+    "no_recorded_attempt": 3,
+    "excluded": 186,
+    "previously_kept": 1323,
+    "restored": 635,
+    "instruct_only_solved_retained": 34
+  }
+}

endless_v2/pass_counts.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a582d131d83d66ae8126f02ede5148cdd06aa1b8d4ace43eb2331a286bcc7e2e
+size 99197

endless_v2/patch.py

@@ -1,21 +1,66 @@
-"""Reproduce the frozen v2 selection from the unchanged, pinned v1 payloads."""
-import argparse,json,hashlib
+"""Reproduce v2 selection from pinned, unchanged v1 payloads and measured pass counts."""
+import argparse
+import hashlib
+import json
 from pathlib import Path
-import pyarrow as pa,pyarrow.parquet as pq
+import pyarrow as pa
+import pyarrow.parquet as pq
 from huggingface_hub import hf_hub_download
-p=argparse.ArgumentParser(description=__doc__);p.add_argument('--out',type=Path,required=True);p.add_argument('--source',type=Path);a=p.parse_args();a.out.mkdir(parents=True,exist_ok=True)
-assert not (a.out/'tasks.parquet').exists()
-m=json.loads((Path(__file__).parent/'metadata/selection.json').read_text());f=a.source or Path(hf_hub_download(m['repo_id'],m['source_folder']+'/tasks.parquet',revision=m['source_revision'],repo_type='dataset',cache_dir=a.out/'.source-cache'));assert hashlib.sha256(f.read_bytes()).hexdigest()==m['source_sha256']
-src=pq.ParquetFile(f);schema=src.schema_arrow.remove_metadata();columns=[model+'_'+metric for model in m['models'] for metric in ['successes','graded_attempts','missing_attempts']]
-for name in columns+['target_attempts_per_model']:schema=schema.append(pa.field(name,pa.int64()))
-rows=[]
-for batch in src.iter_batches(batch_size=32):
- for r in batch.to_pylist():
-  t=m['tasks'][r['path']];assert r['content_sha256']==t['content_sha256']
-  if not t['kept']:continue
-  for k in columns:r[k]=t[k]
-  r['target_attempts_per_model']=4;rows.append(r)
-assert len(rows)==m['summary']['solved'];names={r['path'] for r in rows};im=json.loads(src.schema_arrow.metadata[b'ot.images.v1']);im['tasks']={k:v for k,v in im['tasks'].items() if k in names};bundles={e['bundle'] for v in im['tasks'].values() for e in v['environments']};im['bundles']={k:v for k,v in im['bundles'].items() if k in bundles}
-if 'uncached_task_ids' in im:im['uncached_task_ids']=sorted(names-set(im['tasks']));im['coverage']={'cached_tasks':len(im['tasks']),'total_tasks':len(names)}
-else:assert set(im['tasks'])==names
-meta={b'ot.images.v1':json.dumps(im).encode(),b'ot.release.v2':json.dumps({k:v for k,v in m.items() if k!='tasks'}).encode()};pq.write_table(pa.Table.from_pylist(rows,schema=schema).replace_schema_metadata(meta),a.out/'tasks.parquet',compression='zstd');print('Retained',len(rows),'cached tasks',len(im['tasks']),'bundles',len(bundles))
+
+def main():
+    ap=argparse.ArgumentParser(description=__doc__)
+    ap.add_argument('--out',type=Path,required=True)
+    ap.add_argument('--source',type=Path)
+    args=ap.parse_args()
+    args.out.mkdir(parents=True,exist_ok=True)
+    assert not any((args.out/name).exists() for name in ['tasks.parquet','archive.parquet','pass_counts.parquet'])
+    selection=json.loads((Path(__file__).parent/'metadata/selection.json').read_text())
+    source=args.source or Path(hf_hub_download(selection['repo_id'],selection['source_folder']+'/tasks.parquet',
+        revision=selection['source_revision'],repo_type='dataset',cache_dir=args.out/'.source-cache'))
+    assert hashlib.sha256(source.read_bytes()).hexdigest()==selection['source_sha256']
+    pf=pq.ParquetFile(source)
+    fields=['instruct_successes','instruct_graded_attempts','instruct_missing_attempts',
+        'coder_successes','coder_graded_attempts','coder_missing_attempts',
+        'target_attempts_per_model','instruct_target_attempts','coder_target_attempts']
+    schema=pf.schema_arrow.remove_metadata()
+    for field in fields:schema=schema.append(pa.field(field,pa.int64()))
+    kept=[];excluded=[];counts=[];seen=set()
+    for batch in pf.iter_batches(batch_size=32):
+        for row in batch.to_pylist():
+            task=row['path'];v=selection['tasks'][task]
+            assert task not in seen and row['content_sha256']==v['content_sha256']
+            seen.add(task)
+            retain=v['instruct_successes']>0 or v['coder_successes']>0
+            assert retain==v['kept']
+            row.update({field:v[field] for field in fields})
+            if retain:kept.append(row)
+            else:
+                row.update(archive_stage=6,archive_reason=v['exclusion_reason'],archive_labels=[v['exclusion_reason']])
+                excluded.append(row)
+            for model,budget in selection['target_attempts'].items():
+                if v[model+'_graded_attempts']==budget:
+                    counts.append({'path':task,'content_sha256':v['content_sha256'],'model':selection['models'][model],
+                        'attempts':budget,'graded':budget,'solved':v[model+'_successes'],'retained_v2':retain})
+    assert seen==set(selection['tasks'])
+    assert len(kept)==selection['summary']['solved'] and len(excluded)==selection['summary']['excluded']
+    def metadata(rows):
+        image=json.loads(pf.schema_arrow.metadata[b'ot.images.v1']);names={r['path'] for r in rows}
+        image['tasks']={k:v for k,v in image['tasks'].items() if k in names}
+        bundles={e['bundle'] for task in image['tasks'].values() for e in task['environments']}
+        image['bundles']={k:v for k,v in image['bundles'].items() if k in bundles}
+        if 'uncached_task_ids' in image:image['uncached_task_ids']=sorted(names-set(image['tasks']))
+        image['coverage']={'cached_tasks':len(image['tasks']),'total_tasks':len(names)}
+        assert len(image['tasks'])==len(names)
+        return {b'ot.images.v1':json.dumps(image).encode(),
+            b'ot.release.v2':json.dumps({k:v for k,v in selection.items() if k!='tasks'}).encode()}
+    pq.write_table(pa.Table.from_pylist(kept,schema=schema).replace_schema_metadata(metadata(kept)),args.out/'tasks.parquet',compression='zstd')
+    archive_schema=schema
+    for name,kind in [('archive_stage',pa.int64()),('archive_reason',pa.string()),('archive_labels',pa.list_(pa.string()))]:
+        archive_schema=archive_schema.set(archive_schema.get_field_index(name),pa.field(name,kind))
+    pq.write_table(pa.Table.from_pylist(excluded,schema=archive_schema).replace_schema_metadata(metadata(excluded)),args.out/'archive.parquet',compression='zstd')
+    # Preserve model-first order used in the release pass-count table.
+    counts.sort(key=lambda r:(list(selection['models'].values()).index(r['model']),r['path']))
+    pq.write_table(pa.Table.from_pylist(counts),args.out/'pass_counts.parquet',compression='zstd')
+    print(json.dumps({'kept':len(kept),'excluded':len(excluded),'pass_count_rows':len(counts)}))
+
+if __name__=='__main__':main()

endless_v2/tasks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:977e0ea5376dd03d056c539eed124034d2cfa634d9d8b04e35c3189e751af413
-size 17100121
+oid sha256:7bd559ec4d60290cf09cda5604b232bc144b20d3734a8621efa031153f99a792
+size 28311796

endless_v2/traces/attempts.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a6b6e98e5deaeba08f957e15196d692aa001277e27282f143a10fc53aee8df70
-size 17362864
+oid sha256:63dd8ab34fb39a9c0b50ca4f1246b4c21568564cb31c5e7b356d661213847544
+size 76829713

endless_v2/traces/manifest.json

@@ -1,8 +1,47 @@
 {
-  "attempts": 8564,
+  "attempts": 25692,
   "tasks": 2141,
-  "successful_attempts": 3659,
-  "missing_conversations": 0,
-  "missing_raw_verifier_rewards": 5,
-  "results_revision": "b23edb5a7b3e2989bb16dd5bc337e02399aee7d5"
+  "models": {
+    "instruct": {
+      "model": "Qwen/Qwen3-30B-A3B-Instruct-2507",
+      "tasks": 2141,
+      "attempts_per_task": 4,
+      "attempts": 8564,
+      "successful_attempts": 3659,
+      "solved_tasks": 1323,
+      "zero_success_tasks": 818,
+      "pass_rates": {
+        "pass@1": 0.42725361980383,
+        "pass@2": 0.5309045617312782,
+        "pass@3": 0.5843063988790285,
+        "pass@4": 0.6179355441382531
+      },
+      "missing_conversations": 0,
+      "missing_raw_verifier_rewards": 5
+    },
+    "coder": {
+      "model": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+      "tasks": 2141,
+      "attempts_per_task": 8,
+      "attempts": 17128,
+      "successful_attempts": 11371,
+      "solved_tasks": 1924,
+      "zero_success_tasks": 217,
+      "pass_rates": {
+        "pass@1": 0.6638836992059786,
+        "pass@2": 0.7814439180623207,
+        "pass@3": 0.8285180489757791,
+        "pass@4": 0.8543204110228865,
+        "pass@8": 0.8986454927603923
+      },
+      "missing_conversations": 1,
+      "missing_raw_verifier_rewards": 10
+    }
+  },
+  "successful_attempts": 15030,
+  "missing_conversations": 1,
+  "missing_raw_verifier_rewards": 15,
+  "scope": "All successful and failed attempts for kept and excluded measured tasks; result records and raw verifier rewards are preserved.",
+  "bytes": 76829713,
+  "sha256": "63dd8ab34fb39a9c0b50ca4f1246b4c21568564cb31c5e7b356d661213847544"
 }