datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
patchpilot-patchgen
PatchPilot patch-generation dataset
Supervised fine-tuning chats for PatchPilot's patch generator (A2). Each chat is exactly the
prompt PatchPilot's agent sends to its model, followed by the developers' real fix written in the
agent's SEARCH/REPLACE edit format.
Source
Built by scripts/build_patchgen_data.py (seed 42) from the SWE-bench training split
(princeton-nlp/SWE-bench, train) and the gold files' contents from princeton-nlp/SWE-bench_oracle.
The same seeded… See the full description on the dataset page: https://huggingface.co/datasets/Tejaswiniprabhakaran19/patchpilot-patchgen.patchpilot-localization
PatchPilot fault-localization dataset
Training data for PatchPilot's fault-localization cross-encoder (A1): pairs of a bug report
and a code candidate (a file or a function), labelled 1 if the developers' fix edits that
candidate.
Source
Built by scripts/build_localization_data.py (seed 42) from three public datasets by Princeton NLP:
Dataset
Used for
princeton-nlp/SWE-bench (train split)
issue text, gold patch, repository… See the full description on the dataset page: https://huggingface.co/datasets/Tejaswiniprabhakaran19/patchpilot-localization.sn107-round-bams