Datasets:
| license: other | |
| task_categories: | |
| - text-generation | |
| tags: | |
| - agentic | |
| - code | |
| - rejection-sampling | |
| # recursive-v4-flash-rejected-samples (filtered, 0.7 threshold) | |
| A quality-filtered subset of `RedMod/recursive-v4-flash-rejected-samples`. These | |
| are deepseek-v4-flash trajectories that FAILED their automated tests, but were | |
| kept because an LLM grader judged the attempt substantively close to correct. | |
| ## Filter | |
| Each trace was graded (facts-extraction rubric) into per-trace counts. A trace is | |
| kept iff **all** hold: | |
| - `artifacts_requested == artifacts_delivered` (the agent produced every requested deliverable), and | |
| - `artifacts_delivered > 0`, and | |
| - `artifacts_substantively_matching / artifacts_delivered >= 0.7`. | |
| **46,731 / 86,470 traces (54.0%)** pass. Each row carries its `row_uuid` | |
| (per-trace key). Grades were produced with `openai/gpt-5.6-luna`, which is a | |
| LENIENT grader relative to a stricter baseline (e.g. claude-haiku-4.5) — treat | |
| this subset as a high-recall triage set, not a verified-positive set. | |