Yashp2003's picture
Update logbook: Reproduction: SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?
bd3727b verified
|
Raw
History Blame
4.36 kB

Conclusion


Conclusion

Overall verdict: 5/5 claims structurally corroborated

All five abstract-level claims are consistent with what the public data shows. The benchmark design, split arithmetic, multi-file patch rates, human-augmentation fields, and domain spread all match the paper's description.

What we could verify

Claim What the paper says What we found Verdict
1: Scale 1,865 problems / 41 repos 731 public / 11 repos (11+12+18=41 consistent) partial
2: Split 11 public / 12 held-out / 18 commercial 11 public repos verified; 12+18 locked consistent
3: Long-horizon Hours-to-days, multi-file 85.6% multi-file, median 4 files, mean 5.07 supported
4: Human-verified Adequate context 100% carry requirements/interface/test_patch supported
5: Contamination-resistant Spans business/B2B/dev-tools 4 languages, containerised, post-cutoff supported

What we could NOT verify

  • The full 1,865 / 41 count (held-out and commercial splits are deliberately non-public)
  • Agent solve rates on the benchmark (this was a dataset audit, not an agentic re-solving)
  • Contamination resistance in practice (structural claim only, not a run-time test)

Reproduction bundle

All artifacts are on Hugging Face:

How to rerun

pip install datasets huggingface_hub
python analyze.py

The script loads ScaleAI/SWE-bench_Pro, counts instances and repos, measures patch complexity, checks field coverage, and prints all verification results. No GPU required. Full run takes ~3 minutes on a free CPU tier.


Reproduction Bundle

Primary artifacts (Hugging Face):

Reproduction poster:

How to rerun:

pip install datasets huggingface_hub
python analyze.py