TberiusJunyao/T1-Terminal-Bench-2.1-Trajectories

T1 — Terminal-Bench 2.1 Trajectories

Complete agent trajectories from evaluating T1 (a 122B-parameter Qwen3.5-based model) on Terminal-Bench 2.1 with the Terminus-2 agent scaffold.

Two independent repetitions of the same 89-task suite under an identical frozen config are included, so the run-to-run variance is visible rather than hidden behind a single number.

Both means reconcile exactly with each run's own Harbor result.json stats block. The mean divides by all 89 trials, not by the scored ones — an errored trial enters the average as a zero, with performance of 64.0%.

Layout

data/<run>/<task-name>__<trial-id>/
├── metadata.json                              # reward, timings, token counts, sampling config
├── trajectory.json                            # ATIF-v1.7 — the full agent/user step sequence
├── trajectory.cont-N.json                     # continuation segments after context compaction
├── trajectory.summarization-N-questions.json  # compaction: what the agent asked itself
├── trajectory.summarization-N-answers.json    # compaction: answers used to build the summary
├── trajectory.summarization-N-summary.json    # compaction: the resulting carried-forward summary
├── recording.cast                             # asciinema terminal recording
└── terminus_2.pane                            # final tmux pane capture

metadata.jsonl in the repo root is one flat row per trial (reward, task, token counts, exception type, path) so the dataset is browsable in the HF viewer.

Reading a trajectory

import json
t = json.load(open("data/run_2/adaptive-rejection-sampler__z8NLa2T/trajectory.json"))
print(t["schema_version"], len(t["steps"]))          # ATIF-v1.7 130
for s in t["steps"][:4]:
    print(s["step_id"], s["source"], s["message"][:100])

Each step carries step_id, timestamp, source (user = harness/terminal output, assistant = model), and message. The agent emits JSON containing analysis, plan, and a commands array of keystrokes; the user steps carry the resulting terminal output.

Context compaction

Terminus-2 compacts context when it approaches the input limit (86,016 tokens here). A compaction writes a summarization-N-* triple and opens a new trajectory.cont-N.json segment whose step_id restarts at 1. To reconstruct one episode in order, read trajectory.json then cont-1, cont-2, … in sequence; continued_trajectory_ref links each segment to its successor. Counting steps in trajectory.json alone will undercount episodes that compacted.

Evaluation config

• Agent: Terminus-2 v2.0.0, JSON parser, max_turns 500
• Model: q35-122b-warm2-iter79, served locally via SGLang
• Sampling: temperature 0.1, top_p 0.95, max_tokens 32768, thinking disabled
• Limits: 86,016 input tokens, agent timeout 18,000 s per trial
• Environment: one ephemeral Daytona cloud sandbox per trial

Sanitization

These are raw agent runs, published with infrastructure details removed:

• Removed entirely — internal cluster filesystem paths, job/trial directory URIs and internal job names, the internal inference endpoint, corporate proxy hostnames and IPs, and the cloud organization identifier.
• Dropped files — debug.json (litellm-internal dumps: ~15 GB of re-serialized history, plus internal endpoint and API-key hashes), prompt.txt/response.txt (duplicated by the trajectory steps), and the Harbor result.json/config.json/trial.log/job.log/lock.json bookkeeping files. Every reward and metric was extracted into metadata.json first, so no scores were lost.
• Deliberately kept — /root/..., 127.0.0.1, and Daytona sandbox IDs that appear inside trajectories. These describe the agent's own ephemeral sandbox and are genuine task content.
• One upstream secret retained on purpose — the sanitize-git-repo task ships a planted HuggingFace token in its public fixture repo (github.com/jeffreywpli/test-secret-removal); finding it is the task, so redacting it would corrupt the trajectory. It is not a live credential.

Caveats

• Reward semantics. Terminal-Bench 2.1 rewards here are binary per task (0.0 or 1.0); mean is the solve rate.
• Noise floor is large. Repeat runs of an identical config on this suite spread by roughly 12 percentage points. run_1 and run_2 both landing on 0.6404 is a coincidence of the mean, not evidence of determinism — the per-task pass sets differ. Compare the reward field per task across runs rather than comparing headline numbers.
• Infrastructure errors are not model failures. DaytonaInternalServerError, DaytonaNotFoundError, and similar reflect sandbox-provider issues. AgentTimeoutError means the 5-hour cap was hit.

Citation

@misc{t1_terminal_bench_21_trajectories,
  title  = {T1 Terminal-Bench 2.1 Trajectories},
  author = {Yang, Junyao},
  year   = {2026},
  url    = {https://huggingface.co/datasets/TberiusJunyao/T1-Terminal-Bench-2.1-Trajectories}
}

Model: TberiusJunyao/T1