SWE-chat: Coding Agent Interactions From Real Users in the Wild

Paper: SWE-chat: Coding Agent Interactions From Real Users in the Wild (COLM 2026)

Dataset: SALT-NLP/SWE-chat on HuggingFace

Website: swe-chat.com

---

SWE-chat is a continually growing dataset of real coding agent sessions collected from open-source developers in the wild. The v2 snapshot (data through September, 2026) contains 17,819 sessions across 707 public GitHub repositories, comprising 229,909 user prompts and 2,028,951 agent tool calls, with human vs. agent code authorship attribution.

Demo of data analyses

analyses_and_figures.ipynb is a short tour of the dataset. It contains statistics and figures, starting from the released dataset, which it downloads from HuggingFace. First, request access on the dataset page and log in with a Hugging Face token from an account with access.

pip install -r requirements.txt
hf auth login
jupyter lab analyses_and_figures.ipynb

The notebook needs about 4 GB of working memory; allow additional headroom and roughly 7 GB of disk space for the downloaded tables. For local data, set SWECHAT_LOCAL_DIR to the directory containing the parquet files.

Citation

@inproceedings{baumann2026swechat,
  title     = {SWE-chat: Coding Agent Interactions From Real Users in the Wild},
  author    = {Baumann, Joachim and Padmakumar, Vishakh and Li, Xiang and
               Yang, John and Yang, Diyi and Koyejo, Sanmi},
  booktitle = {Third Conference on Language Modeling},
  year      = {2026},
  url       = {https://arxiv.org/abs/2604.20779}
}