laion/terminal_bench_2_tasktrove_dq_taco_step15_30b_a3b_20260729_222705
Agent trace dataset OpenCode/Harbor rollout traces from the MarinSkyRL run rl-tasktrove-dq-sweep-30b-qwen3-coder-30-20260726-235656-574ba8, exported with make_and_upload_trace_dataset --episodes last (the last episode of each trial — the rollouts the policy was trained on). Coverage Built from the complete trial set on durable object storage, not from a local evidence bundle. quantity value trial directories on object storage 21711 trials with a… See the full description on the dataset page: https://huggingface.co/datasets/laion/terminal_bench_2_tasktrove_dq_taco_step15_30b_a3b_20260729_222705.
Agent trace dataset
OpenCode/Harbor rollout traces from the MarinSkyRL run rl-tasktrove-dq-sweep-30b-qwen3-coder-30-20260726-235656-574ba8, exported with make_and_upload_trace_dataset --episodes last (the last episode of each trial — the rollouts the policy was trained on).
Coverage
Built from the complete trial set on durable object storage, not from a local evidence bundle.
Trials without a result.json were cut off mid-episode and never scored: they carry no reward and no verifier output, so they are excluded.
Secret scan
Every shard was scanned before publication, over decoded string cells rather than the compressed parquet bytes, for JWTs, AWS access-key ids, sk- / hf_ / gh*_ tokens, and PEM private-key markers.
The Iris proxy token that authenticates rollouts to the inference endpoint appears in zero shards. That is structural, not luck: the exporter reads only agent/trajectory.json, the task instruction, the reward, and verifier output. It never reads config.json, exception.txt, agent/opencode.txt, or the full result.json — the result column holds just the reward string. The token is written into config.json and result.json on the source object store, and neither file is an export input.
Four rows contain secret-shaped strings. Each was inspected individually; none is a credential, and all four are agent-transcript or task-prompt content that is kept deliberately — masking it would corrupt the research artifact to satisfy a pattern match.