CrossCodeEval: complete pass@16 results and combined trace statistics
Adds per-task teacher results for all 6,710 CrossCodeEval tasks, with 16 attempts per task for each model (214,720 attempts total).
| Model | Scored attempts | Pass@1 | Pass@16 |
|---|---|---|---|
| Qwen3-30B-A3B-Instruct-2507 | 107,360 | 2.80% | 9.55% |
| Qwen3-30B-A3B (Thinking) | 107,360 | 2.69% | 9.34% |
Includes the original collection and recovery runs, with each counted attempt included once.
- Each language folder has
pass_counts.parquet: one row per task and model, with solved counts and pass@k for every k from 1 through 16. - One combined stage 7 report contains statistics for both models and every task: tokens, turns, runtime, tool use, termination reasons and errors.
- Plots show pass rates for every k from 1 through 16, by language, solved-count distributions and trace statistics. Per-attempt metrics and supporting reports are included.
One attempt has no saved trajectory, so its trace measurements remain missing. One context-limit failure counts as reward 0, with its original error and absent verifier score preserved. Global model-call latency quantiles are unavailable because only per-attempt summaries were retained.
Results and plots: runs/2026-10-07-crosscodeeval-pass16/. The combined statistics are in stage-7-combined.json.gz; summary.json provides a smaller overview.
Trajectories (added later): runs/2026-10-07-crosscodeeval-pass16/trajectories/<model>/<language>.parquet, one row per counted attempt (214720 rows, 214 MB): the agent's trajectory.json, the trial's result.json with the verifier output, the reward and the source submission. 1 attempt(s) have no saved trajectory (trajectories/manifest.json lists them and the truncated evidence archives).