terminal_bench_2_tasktrove_dq_unix_step10_30b_a3b_20260730_014756

# terminal_bench_2_tasktrove_dq_unix_step10_30b_a3b OpenCode agent trajectories from the TaskTrove DQ `unix` arm of a Qwen3-Coder-30B-A3B agentic RL sweep, exported from the complete Harbor rollout artifact set. ## Coverage Built from the full `trace_jobs` prefix of run `rl-tasktrove-dq-sweep-30b-qwen3-coder-30-20260727-082204-e42f1d` (12034 trial directories, 11937 of them scored). | quantity | value | | --- | ---: | | scored trials (`result.json`) | 11937 | | rows published | 11937 | | coverage | 100.00% | Rows are counted against scored trials, not trial directories. The 97 trials without a `result.json` were cut off mid-episode and never scored — they have no reward and no verifier output, so they are excluded by construction. A previous version of this dataset held 248 rows (2.1%) because it was built from a local evidence bundle that mirrors only the most recently modified traces. It has been replaced. ## Fields One row per episode. `conversations` is the ShareGPT-style message list, `instruction` the task prompt, `result` the trial outcome, and `verifier_output` the grader's output where present. ## Redaction One terminal observation contained an RSA private key that the agent generated inside its own ephemeral sandbox during a certificate-generation task. The key block is replaced with `[REDACTED PRIVATE KEY]`. No rows were dropped.