Include SWE LEGO rerun: 1245 kept tasks, 96 new traces and refreshed statistics

swelego_v2/README.md

@@ -1,34 +1,39 @@
 # SWE-Lego v2
 
-Results snapshot: **2026-10-18T21:28:57.295598+00:00**.
-
-## Published dataset stats
+## Dataset stats
 
 - **v1 tasks:** 4,277
-- **Kept:** 1,240 (solved at least once)
-- **Excluded:** 3,037 (no recorded success in this published snapshot)
-- **Tasks with recorded attempts:** 4,265 / 4,277
-- **Recorded attempts:** 34,120
-- **Available conversations:** 18,582
+- **Kept:** 1,245 (solved at least once)
+- **Not selected:** 3,032 (no recorded success)
+- **Newly kept in this update:** 5
+- **Tasks with recorded attempts:** 4,277 / 4,277
+- **Recorded attempts:** 34,216
+- **Available conversations:** 18,678
 - **Missing conversations:** 15,538
 - **Unknown attempt outcomes:** 1,331 across 450 tasks
-- **Tasks with all eight classified outcomes:** 3,815 / 4,277
+- **Tasks with all eight classified outcomes:** 3,827 / 4,277
+- **Task payload changes:** 0
+- **v1 base revision:** `819fe8c3809d9a393502017055c21057b21f02fd`
+
+Results snapshot: **2026-10-20T19:39:53.280360+00:00**. This release selects tasks with at least one observed teacher success and preserves the pinned v1 task payloads.
 
 ## Pass@8 collection
 
-| Model | Tasks attempted | Recorded attempts | Tasks with all 8 outcomes | Tasks solved |
+| Model | Tasks attempted | Recorded attempts | Tasks with all 8 classified outcomes | Tasks solved |
 | --- | ---: | ---: | ---: | ---: |
-| Qwen3-30B Instruct | 4,265 / 4,277 | 34,120 | 3,815 / 4,277 | 1,240 |
+| Qwen3-30B Instruct | 4,277 / 4,277 | 34,216 | 3,827 / 4,277 | 1,245 |
 
-The table describes the files currently published in this folder. It does not yet provide complete dataset-wide pass@8 results.
+The teacher is `Qwen/Qwen3-30B-A3B-Instruct-2507`, using Terminus-2 and eight attempts per task. The completed 12-task Helma rerun is now included: **96 attempts with saved trajectories and five newly solved tasks**. One rerun attempt ended with `TmuxSessionEndedError`; its raw verifier grade remains missing and its outcome reward is zero under the terminal-error policy. [Rerun selection](metadata/rerun-selection-20261020.json) and [stage-6/7 reports and contract](metadata/rerun-reports-20261020.tar.gz) preserve the evidence.
 
-**Rerun update (2026-10-20):** The remaining 12 tasks finished on Helma on October 19. All 96 attempts have saved trajectories; five tasks were solved at least once. One attempt ended with `TmuxSessionEndedError`, has no verifier grade, and scores zero under the recorded terminal-error policy. These rerun results are verified locally but are **not yet included in the published task table, traces, or plots**. Incorporating them would bring the kept count to **1,245** and recorded attempts to **34,216**.
+The rerun preserves temperature 0.7, top-p 0.8, top-k 20, a 32,768-token context and an 8,192-token maximum response. Stage-6 and stage-7 reports are complete for all 12 rerun tasks. These teacher stages do not establish a new full datasource-validation result.
+
+Dataset-wide pass@k estimates remain unavailable because 1,331 historical outcomes are unknown. They are not counted as failures.
 
 ### Successful attempts among kept tasks
 
 | Successful attempts | Instruct |
 | --- | ---: |
-| 1/8 | 463 |
+| 1/8 | 468 |
 | 2/8 | 221 |
 | 3/8 | 144 |
 | 4/8 | 133 |
@@ -37,34 +42,35 @@ The table describes the files currently published in this folder. It does not ye
 | 7/8 | 58 |
 | 8/8 | 72 |
 
-## Reproduce and select tasks
+Counts reflect recorded successes; unknown historical outcomes remain unknown.
 
-`patch.py` rebuilds `tasks.parquet` from v1 at revision `819fe8c3809d9a393502017055c21057b21f02fd` and the accompanying `metadata/selection.json`. The pinned source SHA-256 is verified. Task payloads are current v1 payloads; traces retain the version hashes on which each attempt was evaluated.
+## Evidence and trajectories
 
-31 kept tasks have traces from earlier task versions. The downloadable Parquet contains the current v1 instructions.
+The combined `traces/` files contain **34,216 unique attempt records**, including kept and unselected tasks. **18,678** have saved conversations. Original result JSON is preserved where available, together with raw verifier rewards, outcome rewards, task-version hashes and the current selection flag.
 
-```python
-import pyarrow.parquet as pq
-subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])
-```
+### Missing historical evidence in the HF files
 
-## Evidence and trajectories
+- **15,538 attempt rows lack a saved conversation.** Their attempt records are present. A missing conversation does not by itself mean that the verifier outcome is unknown.
+- **1,580 attempts lack a raw verifier grade.** Of these, **249** have classified terminal-error outcomes and count as reward zero. The other **1,331** have unknown outcomes and no original result JSON in the trace rows; they remain unknown.
+- **450 tasks lack eight classified outcomes** because of these historical unknowns. All 4,277 tasks now have eight attempt records; the 12-task rerun is no longer pending.
 
-The combined `traces/` files contain **34,120 unique attempts** from kept and excluded tasks, including the newer recovery results. Saved conversations are available for **18,582** attempts. Original result JSON is preserved where available, alongside the raw verifier reward, outcome reward, and current `retained_v2` flag.
+These gaps overlap and should not be added together. Recovery from original run archives remains pending; empty fields do not establish that the original traces were lost.
 
-### Missing evidence in the current HF files
+[Trace statistics](trace_stats/README.md) and plots have been regenerated from this published snapshot, including the 96 rerun attempts. Missing measurements are kept separate from measured outcomes.
+
+## Reproduce and select tasks
 
-- **15,538 attempt rows lack a saved conversation.** Their attempt records are present; a missing conversation does not by itself mean that the verifier outcome is unknown.
-- **1,579 attempts lack a raw verifier grade.** Of these, **248** already have recorded terminal-error evidence and count as reward zero. The other **1,331** have unknown outcomes and no original result JSON in the published trace rows. They remain unknown, rather than being assigned zero without evidence.
-- **12 tasks have no attempt rows in this published snapshot.** Their completed rerun results are described above and await incorporation.
+`patch.py` rebuilds `tasks.parquet` from v1 at revision `819fe8c3809d9a393502017055c21057b21f02fd` and `metadata/selection.json`, verifying the pinned source SHA-256. Task payloads are unchanged. Traces identify the version evaluated; **31 kept tasks** have traces from earlier task versions.
 
-The 1,331 unknown outcomes affect 450 tasks. Together with the 12 rerun tasks not yet incorporated, this leaves 462 tasks without eight classified outcomes in the published snapshot. The historical evidence gaps overlap and should not be added together. Recovery from the original run archives remains pending; the empty fields are not proof that those original traces were lost.
+```python
+import pyarrow.parquet as pq
+subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])
+```
 
-[Trace statistics](trace_stats/README.md) and plots describe the existing published traces, including their missing evidence. They have not yet been regenerated with the 96 rerun attempts.
+The `instruct_graded_attempts` column counts classified outcomes, including recorded terminal errors scored zero. `instruct_verifier_graded_attempts` counts actual verifier grades, and `instruct_missing_attempts` counts unknown outcomes out of eight.
 
 ## Runtime
 
-- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**.
-- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches).
+The Helma rerun used Apptainer **1.5.4**, Python **3.12.14** and Harbor **0.8.1**, with [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) and [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). Measured token usage, commands and runtime distributions are in the trace statistics.
 
 Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution).

swelego_v2/metadata/partial-refresh-20261020.json

@@ -0,0 +1,19 @@
+{
+  "baseline_revision": "31605242e5cfb104b8c28c7b28521297db65e291",
+  "snapshot_utc": "2026-10-20T19:39:53.280360+00:00",
+  "summary": {
+    "instruct": {
+      "attempts": 34216,
+      "graded": 32885,
+      "raw_graded": 32636,
+      "complete": 3827,
+      "evaluated": 4277,
+      "solved": 1245
+    },
+    "tasks": 4277,
+    "kept": 1245,
+    "previous_kept": 1240
+  },
+  "policy": "Recorded terminal errors score zero; raw verifier rewards are preserved. Historical unknown outcomes remain unknown.",
+  "trace_policy": "All previous selected attempt IDs plus the 96 attempts from the completed 12-task rerun. No failed submissions or duplicate race copies are added."
+}

swelego_v2/metadata/rerun-reports-20261020.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18de9568b31fc4d58e3214e0dd22a4cd915231d4b8e0d6a30fbad1ae050a8231
+size 114927

swelego_v2/metadata/rerun-selection-20261020.json

@@ -0,0 +1,41 @@
+[
+  {
+    "shard": 18,
+    "stage6": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0018/report/stage-6-36e9fdf426ba.json",
+    "stage7": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0018/report/stage-7-c94ab0464002.json",
+    "contract_sha256": "927e4ea958217477f482306d6d69e1de24712fa3670ea5b685fbf22a22628937",
+    "tasks": [
+      "DigitalSlideArchive__tifftools-69",
+      "deshima-dev__deshima-sensitivity-76",
+      "matplotlib__matplotlib-18446",
+      "scrapy__scrapy-5384"
+    ],
+    "attempts": 32
+  },
+  {
+    "shard": 19,
+    "stage6": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0019/report/stage-6-cfc5fe5620d2.json",
+    "stage7": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0019/report/stage-7-3c5dadc2ae00.json",
+    "contract_sha256": "927e4ea958217477f482306d6d69e1de24712fa3670ea5b685fbf22a22628937",
+    "tasks": [
+      "Dorthu__openapi3-30",
+      "deshima-dev__deshima-sensitivity-81",
+      "matplotlib__matplotlib-18540",
+      "scrapy__scrapy-5482"
+    ],
+    "attempts": 32
+  },
+  {
+    "shard": 96,
+    "stage6": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0096/report/stage-6-0b92535552c2.json",
+    "stage7": "/hnvme/workspace/y500bb12-optiagent/runs/swelego-helma-20261016/full/array/shards/0096/report/stage-7-52d7d70f5ecb.json",
+    "contract_sha256": "927e4ea958217477f482306d6d69e1de24712fa3670ea5b685fbf22a22628937",
+    "tasks": [
+      "audeering__auglib-24",
+      "iterative__dvc-4075",
+      "pymor__pymor-1296",
+      "tornadoweb__tornado-2075"
+    ],
+    "attempts": 32
+  }
+]

swelego_v2/metadata/verification-20261020.json

@@ -0,0 +1,10 @@
+{
+  "unique_attempts": 34216,
+  "kept": 1245,
+  "source_payloads_unchanged": true,
+  "prior_outcomes_unchanged": true,
+  "reproduction_identical": true,
+  "missing_trajectories": 15538,
+  "raw_missing": 1580,
+  "unknown": 1331
+}

swelego_v2/tasks.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48190112ae7d847b16ece3b180d78e4ce074bc6aee6c2cd6f690a1bb97b691bf
-size 5595470
+oid sha256:095268bf849c864428e9fa3922f6243051cbb7e490e8a369a3d8b5394f51364a
+size 5621360

swelego_v2/trace_stats/README.md

@@ -1,13 +1,13 @@
 # Trajectory statistics
 
-34,120 attempts across 4,265 tasks of SWE LEGO v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md).
+34,216 attempts across 4,277 tasks of SWE LEGO v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md).
 
 | Metric | instruct |
 | --- | ---: |
-| Recorded attempts | 34,120 |
+| Recorded attempts | 34,216 |
 | Median agent turns | 15 |
-| Median input tokens per attempt | 96,305 |
-| Median output tokens per attempt | 4,806 |
+| Median input tokens per attempt | 96,423 |
+| Median output tokens per attempt | 4,814 |
 | Median agent runtime (seconds) | 101 |
 | Median verifier runtime (seconds) | 9 |
 
@@ -31,4 +31,4 @@ Per-model plots: [instruct](plots/instruct-trace-commands.png), [instruct](plots
 - [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON.
 - [Provenance](data/provenance.json): source files and deduplication.
 
-Attempts without a saved conversation: 15,538; without a verifier reward: 1,579. Their missing measurements stay empty.
+Attempts without a saved conversation: 15,538; without a verifier reward: 1,580. Their missing measurements stay empty.

swelego_v2/trace_stats/data/per_attempt.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:501077608a522cecfb0448321437e50341f543845976cbfaffd6d035d1bc8060
-size 3966497
+oid sha256:4811265ee47cf5dd1f4ef74fd1eec482452d4d8d9b1fa5b7e95799744c073103
+size 3982459

swelego_v2/trace_stats/data/per_task.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bb398eb642a2b8969a1d2ea19a9835fc50e864ef0e75f4e857335d08c4637695
-size 1360470
+oid sha256:16296906f68af06b4b4a2802e051321dd4abfbb8fddbd4c8b0b3dcc888fd9fcd
+size 1368772

swelego_v2/trace_stats/data/provenance.json

@@ -1607,6 +1607,12 @@
    "rows": 72,
    "unique_rows_added": 72,
    "sha256": "dcfbd54a353f70966448b9cac5f4f110f9820d6d77e918e0ecfb3ba6c6678e16"
+  },
+  {
+   "file": "traces/part-0267.parquet",
+   "rows": 96,
+   "unique_rows_added": 96,
+   "sha256": "f67486acb73a42ba5a4961db60bf9905fd7a024d37d97f6f2117dfb6b3f9e82c"
   }
  ],
  "figures": {

swelego_v2/traces/manifest.json

@@ -1,12 +1,12 @@
 {
-  "snapshot_utc": "2026-10-18T21:28:57.295598+00:00",
+  "snapshot_utc": "2026-10-20T19:39:53.280360+00:00",
   "counts": {
-    "attempts": 34120,
+    "attempts": 34216,
     "missing_trajectories": 15538,
-    "raw_missing": 1579,
-    "terminal_errors_zero": 248,
+    "raw_missing": 1580,
+    "terminal_errors_zero": 249,
     "unknown_outcomes": 1331
   },
   "reward_policy": "terminal-errors-zero-v2",
-  "selection": "Combined unique attempts for the pass8 collection, including the recovered Capella attempts. Task version hashes identify the evaluated payloads."
+  "selection": "All previous selected attempt IDs plus the 96 attempts from the completed 12-task rerun. No failed submissions or duplicate race copies are added."
 }

swelego_v2/traces/part-0267.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f67486acb73a42ba5a4961db60bf9905fd7a024d37d97f6f2117dfb6b3f9e82c
+size 1300209