Focus BugsInPy v3 READMEs on the new pass@2 evaluation

bugsinpy_v3/README.md

@@ -2,12 +2,11 @@
 
 ## Dataset stats
 
-- **Tasks evaluated across all models:** 489
+- **Source tasks:** 489
 - **Kept:** 471 (65 from v2 plus 406 newly solved by Qwen3.8-27B)
 - **Excluded:** 18 (no recorded success)
 - **New evaluation:** 424 previously excluded tasks, 2 attempts each, 848 attempts
-- **All recorded attempts:** 16,496, including 15,648 from v2
-- **Saved trajectories:** 16,496, successful and failed
+- **New saved trajectories:** 848, successful and failed
 
 BugsInPy contains Python bug-fixing tasks from real open-source projects. Each task provides a buggy project checkout and a verifier for the target bug. `tasks.parquet` keeps all v2 tasks and adds the newly solved tasks with their evaluated payloads.
 
@@ -15,11 +14,9 @@ BugsInPy contains Python bug-fixing tasks from real open-source projects. Each t
 
 | Model | Tasks evaluated | Attempts | Tasks solved at least once | Pass@1 | Pass@k |
 | --- | ---: | ---: | ---: | ---: | ---: |
-| Qwen3-30B Instruct | 489 | 7,824 | 56 | 3.13% | 11.45% (k=16) |
-| Qwen3-30B Thinking | 489 | 7,824 | 47 | 2.28% | 9.61% (k=16) |
 | Qwen3.8-27B (thinking) | 424 | 848 | 406 | 90.92% | 95.75% (k=2) |
 
-The two Qwen3-30B models solved 65 unique tasks together. Qwen3.8-27B evaluated the other 424 tasks and added 406, bringing the kept set to 471 of 489 tasks (96.32%). The 65 v2 tasks were not reevaluated with Qwen3.8-27B.
+Qwen3.8-27B solved 406 of the 424 previously excluded tasks. Together with the 65 tasks carried forward from [v2](../bugsinpy_v2/README.md), v3 keeps 471 tasks. These results cover only the new evaluation.
 
 Qwen3.8-27B used Terminus-2, thinking enabled, temperature 1.0, top-p 0.95, top-k 20, a 32,768-token context, and at most 8,192 output tokens per model call. Each task kept its existing agent and verifier time limits. The task source is pinned to revision `3abfc632`; complete settings are in [evaluation setup](metadata/evaluation-setup.json).
 
@@ -33,30 +30,6 @@ Qwen3.8-27B used Terminus-2, thinking enabled, temperature 1.0, top-p 0.95, top-
 
 ![Qwen3.8-27B success distribution](trace_stats/plots/qwen38-success-counts.png)
 
-### Successful attempts in the earlier evaluations
-
-Counts below cover all 489 evaluated tasks, including those excluded from v2.
-
-| Successful attempts | Instruct tasks | Thinking tasks |
-| --- | ---: | ---: |
-| 0/16 | 433 | 442 |
-| 1/16 | 16 | 17 |
-| 2/16 | 8 | 7 |
-| 3/16 | 7 | 4 |
-| 4/16 | 3 | 4 |
-| 5/16 | 6 | 3 |
-| 6/16 | 5 | 1 |
-| 7/16 | 1 | 5 |
-| 8/16 | 0 | 0 |
-| 9/16 | 2 | 2 |
-| 10/16 | 1 | 2 |
-| 11/16 | 3 | 1 |
-| 12/16 | 2 | 0 |
-| 13/16 | 1 | 0 |
-| 14/16 | 0 | 1 |
-| 15/16 | 1 | 0 |
-| 16/16 | 0 | 0 |
-
 ## Select tasks by successful-attempt count
 
 The Parquet preserves `instruct_successes`, `thinking_successes`, and `attempts_per_model` from v2, and adds `qwen38_successes`, `qwen38_attempts`, and `kept_from_v2`. For the 65 v2 tasks, `qwen38_attempts` is 0 and `qwen38_successes` is null.
@@ -74,7 +47,7 @@ v2_tasks = pq.read_table("tasks.parquet", filters=[("kept_from_v2", "=", True)])
 
 ## Traces
 
-`traces/` holds all 16,496 attempts, successful and failed, one row per attempt in Zstandard-compressed Parquet. This includes all v2 traces and all 848 new Qwen3.8-27B traces. Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2`, `retained_v3`, and full `trajectory` and `result` JSON.
+`traces/` includes all 848 new Qwen3.8-27B attempts, successful and failed, one row per attempt in Zstandard-compressed Parquet. Earlier traces are also preserved; use the model filter below to select the new evaluation. Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2`, `retained_v3`, and full `trajectory` and `result` JSON.
 
 ```python
 import json
@@ -99,8 +72,6 @@ Not yet checked with the repository-leak checks added on 2026-10-17 (leftover gi
 
 | Model | Mean time per attempt | Median time per attempt | Attempts measured |
 | --- | ---: | ---: | ---: |
-| Qwen3-30B Instruct | 819.81 s | 591.56 s | 7,824 |
-| Qwen3-30B Thinking | 1,098.50 s | 750.63 s | 7,824 |
 | Qwen3.8-27B | 1,172.06 s | 935.76 s | 848 |
 
 `patch.py` rebuilds `tasks.parquet` from pinned v1 and v2 sources and the published selection metadata.

bugsinpy_v3/trace_stats/README.md

@@ -1,34 +1,32 @@
 # Trajectory statistics
 
-16,496 attempts across 489 tasks of bugsinpy_v3, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md).
+848 new Qwen3.8-27B attempts across 424 evaluated tasks, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md).
 
-| Metric | qwen3-30b-instruct-2507 | qwen3-30b-thinking | qwen3.8-27b |
-| --- | ---: | ---: | ---: |
-| Recorded attempts | 7,824 | 7,824 | 848 |
-| Median agent turns | 35 | 32 | 21 |
-| Median input tokens per attempt | 462,749 | 383,801 | 247,841 |
-| Median output tokens per attempt | 21,119 | 35,764 | 26,467 |
-| Median agent runtime (seconds) | 346 | 500 | 761 |
-| Median verifier runtime (seconds) | 18 | 17 | 34 |
+| Metric | Qwen3.8-27B |
+| --- | ---: |
+| Recorded attempts | 848 |
+| Median agent turns | 21 |
+| Median input tokens per attempt | 247,841 |
+| Median output tokens per attempt | 26,467 |
+| Median agent runtime (seconds) | 761 |
+| Median verifier runtime (seconds) | 34 |
 
 Medians use recorded measurements only.
 
 ## Plots
 
-![Trace measurements](plots/all-models-trace-distributions.png)
+![Trace measurements](plots/qwen3.8-27b-trace-distributions.png)
 
-![Trace outcomes](plots/all-models-trace-outcomes.png)
+![Trace outcomes](plots/qwen3.8-27b-trace-outcomes.png)
 
-![Most frequent commands](plots/all-models-trace-commands.png)
+![Most frequent commands](plots/qwen3.8-27b-trace-commands.png)
 
 Commands are counted as submitted, not as successful executions.
 
-Per-model plots: [qwen3-30b-instruct-2507](plots/qwen3-30b-instruct-2507-trace-commands.png), [qwen3-30b-instruct-2507](plots/qwen3-30b-instruct-2507-trace-distributions.png), [qwen3-30b-instruct-2507](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [qwen3-30b-thinking](plots/qwen3-30b-thinking-trace-commands.png), [qwen3-30b-thinking](plots/qwen3-30b-thinking-trace-distributions.png), [qwen3-30b-thinking](plots/qwen3-30b-thinking-trace-outcomes.png), [qwen3.8-27b](plots/qwen3.8-27b-trace-commands.png), [qwen3.8-27b](plots/qwen3.8-27b-trace-distributions.png), [qwen3.8-27b](plots/qwen3.8-27b-trace-outcomes.png).
-
 ## Data
 
-- [Summary](data/summary.json): overall and per-model statistics.
-- [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.
+The detailed files preserve all evaluations. Select `qwen3.8-27b` in the `model` column of [per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), or use `models["qwen3.8-27b"]` in the [summary](data/summary.json), for the new evaluation. Nested metrics are stored as JSON. [Provenance](data/provenance.json) records source files and deduplication.
+
+Earlier evaluation statistics are in the [v2 overview](../../bugsinpy_v2/trace_stats/README.md).
 
-Attempts without a saved conversation: 0; without a verifier reward: 2,711. Their missing measurements stay empty.
+New attempts without a saved conversation: 0; without a verifier reward: 3. Missing measurements stay empty.