yangjunxiao2021/terminal-bench-science-trajectories

Terminal-Bench Science trajectories

An archival snapshot of public Harbor evaluation runs, with original ATIF trajectories and flat analysis tables.

• TB Science 轨迹分析 — online comparison reader
• Code and download scripts

This is a partial capture, not a claim that all trajectories were downloaded.

Snapshot	Value
Captured at (UTC)	2026-10-17T16:18:53.584574+00:00
Model / agent configurations	17
Tasks	70
Publicly listed trials	2520
Source trajectories captured	2507
Download errors	0
Source returned no trajectory	13
Not downloaded yet	0
Official trials with no public listing	1050

Raw files upload incrementally. upload-status.json records whether the complete prepared snapshot has finished uploading. Capture counts above describe the local snapshot; during upload, a raw object listed in the index may not be on the Hub yet.

Layout

• analysis/trials.csv: one row per listed trial, including missing/error/pending states.
• analysis/models.csv: official leaderboard metrics and archive coverage for every configuration.
• analysis/tasks.csv: per-task counts and observed scored-run success rates.
• trajectories/models/<model>/<configuration-id>/<task>.jsonl.gz: lossless JSONL, one full run per line, including repeated runs. trajectory is the unmodified original ATIF object when valid, or null with an explicit download state. Nonstandard source JSON is preserved exactly in raw_trajectory; a separately labeled reading_trajectory may project unquoted [REDACTED] markers to null.
• manifest.json, catalog.json, inventory.json: provenance and file checksums.

The Dataset Viewer reads only the explicitly configured flat CSV tables. It does not infer an Arrow schema across heterogeneous nested ATIF objects.

Analyze

from datasets import load_dataset
df = load_dataset("yangjunxiao2021/terminal-bench-science-trajectories", "trials", split="test").to_pandas()
scored = df[(df.is_scored == 1) & df.reward.notna()]
rates = scored.assign(passed=scored.reward.eq(1)).groupby("model").passed.mean()
print(rates.sort_values(ascending=False))

import gzip, json
from huggingface_hub import hf_hub_download
row = df.loc[df.download_status.eq("downloaded")].iloc[0]
path = hf_hub_download("yangjunxiao2021/terminal-bench-science-trajectories", row.raw_path, repo_type="dataset")
with gzip.open(path, "rt") as f:
    runs = [json.loads(line) for line in f]
trajectory = next(r["trajectory"] for r in runs if r["trial_id"] == row.trial_id)

Method and provenance

Source: https://hub.harborframework.com/datasets/terminal-bench-science/terminal-bench-science/latest?tab=leaderboard. All source trial links, job IDs, model / agent settings, errors, and retrieval timestamps are preserved. Private, unlisted and deleted source runs are not reconstructed. Missing reward values are not treated as zero. Observed analysis rates use is_scored == 1 and numeric reward; success means reward equals 1. Official leaderboard rates use the source's own denominator and are kept separately. Run ordinals are not matched seeds.

This dataset is intended for analysis of evaluation behavior. Original benchmark canary text and embedded images are preserved. Code licensing is separate from upstream benchmark content; no new license is assigned to that content.