zkjzou/swe-verifier-training

SWE-Verifier training data

Training data for the three trained models used in SWE-Verifier.

Models

Paper variant	Hugging Face model
Score Only	zkjzou/SWE-Verifier-Score-Only
Joint Rubric	zkjzou/SWE-Verifier-Joint-Rubric
Two-Stage Rubric	zkjzou/SWE-Verifier-Two-Stage-Rubric

Data

Configuration	Rows	Split
score_only	30,494	train
joint_rubric	30,403	train
two_stage_generator	30,403	train
two_stage_scorer	30,403	train
raw_teacher	6,669 request/response pairs	corpus

Training rows contain messages, with the final assistant message as the target. The raw teacher pairs cover 33,343 candidate actions. They contain saved teacher inputs and output text; provider metadata is excluded, and detected sensitive patterns are redacted. The redacted field identifies affected rows.

from datasets import load_dataset

training = load_dataset("zkjzou/swe-verifier-training", "score_only", split="train")
raw = load_dataset("zkjzou/swe-verifier-training", "raw_teacher", split="corpus", streaming=True)

Decode input_json with json.loads to read a teacher input. teacher_raw_output contains its output text. Action identifiers link raw records to the training manifests.

The repository is private during release preparation; an authorized Hugging Face login is required.

License

Original annotations and packaging: Apache 2.0. Embedded upstream material retains its original licenses; notices are included in upstream_licenses/.

Paper · Code