zlab-princeton/SWEeper-Bench-traces

🧹 SWEeper-Bench agent traces

Paper · Blog · Code · Benchmark

This dataset has every run behind the SWEeper-Bench paper: 5,200 runs from 42 agent configurations. For each run you get the prompt the coding agent saw, its full trajectory, the shell commands it ran against the app, its patch, and the browser verifier's trajectories and verdicts.

Main results

All 15 agents ran on the 200 tasks with the open-ended prompt and the browser prompt template. A task passes only when both of its behavior tests pass. A task with no prediction or evaluation counts as a fail.

Model	Harness	Effort	Pass rate
grok-4.6	Cursor	xhigh-fast	59.0%
qwen3.8-flash	Codex	xhigh	57.5%
claude-fable-5-1	Claude Code	xhigh	56.5%
claude-opus-5	Claude Code	xhigh	56.5%
gpt-6-astra	Codex	xhigh	56.5%
deepseek-v4.1-flash	DeepSeek harness	xhigh	51.5%
kimi-k3	Kimi Code	high	48.5%
deepseek-v4-pro	DeepSeek harness	xhigh	47.0%
gpt-5.6-sol	Codex	xhigh	47.0%
muse-spark-1.3	Muse Code	xhigh	47.0%
glm-5.3	Codex	max	43.5%
gpt-5.6-luna	Codex	xhigh	42.0%
glm-5.3-flash	Codex	max	39.5%
gpt-5.6-terra	Codex	xhigh	35.0%
qwen3.8-max	Codex	none	28.0%

Ablations

type	What changes	Runs
ablation-prompt	The agent gets a description of the bug (specified prompt). Five agents pass 96.0% to 96.5%.	1,000
ablation-template	The prompt no longer asks the agent to test in a browser (non-browser template).	400
ablation-harness	gpt-5.6-luna in four other harnesses, on a 40-task subset.	160
ablation-time	Time budgets of 20, 40, 60, and 80 minutes, on the same 40 tasks.	640

Finding runs

runs.csv has one row per run, and the dataset viewer shows it as the runs split.

Column	Meaning
run_id	Folder name of the run
task_id	sweeper-001 through sweeper-200
model, harness, effort	The agent
prompt	scoped (open-ended) or specified (bug described)
template	browser or non-browser
set	200 (all tasks) or 40 (ablation subset)
time_budget	unlimited or minutes
type	main or one of the ablations above
target, preservation	Verdict of each behavior test
overall	pass only when both tests pass
hf_url	Link to the run's folder

Download a run

from huggingface_hub import hf_hub_download, snapshot_download
import pandas as pd

repo = "zlab-princeton/SWEeper-Bench-traces"
runs = pd.read_csv(hf_hub_download(repo, "runs.csv", repo_type="dataset"))

run = runs.query("model == 'gpt-6-astra' and task_id == 'sweeper-001' and type == 'main'").iloc[0]
path = snapshot_download(repo, repo_type="dataset", allow_patterns=f"{run.run_id}/*")

The whole dataset is about 5 GB in 65,000 files. To get all of it, git clone https://huggingface.co/datasets/zlab-princeton/SWEeper-Bench-traces is much faster than snapshot_download.

Files in a run

<run_id>/
  metadata.json                 Same fields as the runs.csv row
  prediction/
    prompt.txt                  Prompt given to the coding agent
    harness.stdout.txt          Agent output, including its trajectory
    harness.stderr.txt          Agent error output
    trajectory.jsonl            Claude Code only: stream events
    final.txt                   Claude Code only: final message
    status.shell.txt            Commands the agent ran in the app container
    patch                       The agent's patch
  evaluation/
    result.json                 Verdicts for both behavior tests
    <test-id>/
      prompt.txt                Verifier instructions
      trajectory.jsonl          Verifier trajectory in the browser
      last-message.txt          Verifier's final verdict

Some files are missing when a run stopped before that stage. The hf_url inside each metadata.json points to the original upload; use the one in runs.csv.

License

This dataset is released under CC BY 4.0. Traces quote source code from the benchmark applications, which keep their own licenses.

Citation

@article{yao2026sweeperbench,
  title   = {SWEeper-Bench: Can Agents Discover Bugs in Interactive Software?},
  author  = {Yao, Yang and Chen, Haozhe and Kang, Bingyi and
             Narasimhan, Karthik R and Liu, Zhuang},
  journal = {arXiv preprint arXiv:2610.14096},
  year    = {2026}
}