Datasets:

umich-fatml
/

rollouts

configs:
  - config_name: terminal-bench-2-1
    default: true
    data_files:
      - split: train
        path: data/terminal-bench-2-1/*.parquet
tags:
  - agents
  - harbor
  - terminal-bench

UMich FATML Harbor Rollouts

This private dataset contains Harbor agent rollouts in an extensible, one-row-per-trial format. The terminal-bench-2-1 configuration currently contains 1,780 rollouts over 89 tasks: 749 passing, 1,021 failing, and 10 unscored.

Record format

Frequently filtered identity, timing, token, reward, and status fields are typed Parquet columns. Lossless structured payloads are canonical JSON strings:

trajectory_json: ATIF for the current collection, identified by the adjacent format and schema-version columns.
harness_result_json: the sanitized Harbor result.json payload.
verifier_result_json: Harbor's arbitrary verifier result and reward map.
verifier_artifacts_json: textual verifier artifacts with paths, media types, sizes, SHA-256 digests, and content. Binary artifacts can be represented by an external path.
run_json and run_stats_json: sanitized run-level provenance and aggregate statistics.

Missing trajectory or verifier data is represented with null payloads and explicit presence fields; trials are never dropped solely because an artifact is unavailable. Host-local paths, credentials, and internal endpoint/routing values are removed or normalized. Raw mini-swe-agent provider dumps are intentionally excluded because they repeat rendered prompts and response payloads already represented by ATIF.

Source limitations in this release: 5 trajectories and 11 CTRF files are absent, and 20 existing verifier artifacts are recorded as unavailable because their source permissions do not allow them to be read.

Provenance

The current data uses mini-swe-agent 2.4.5 with k2moe375b-mid4-6000 on Terminal-Bench 2. Terminal-Bench 2 source tasks are distributed under Apache-2.0; this private dataset does not declare a dataset-wide license.