CoolFace
Datasetpublic

LSW142857/OPSD-PI-SWE-Gym-1024

OPSD-PI SWE-Gym 1024 OPSD-PI is a stage-adaptive privileged-information distillation project for repository-level coding agents. The student sees only the ordinary task and interaction history; a frozen teacher additionally receives stage-specific PI for EXPLORE, REPRODUCE, DIAGNOSE, EDIT, or VERIFY. Top-K forward KL distills that signal into the student LoRA and MTP head, while evaluation runs without external PI. This repository publishes the prepared teacher-side inputs and… See the full description on the dataset page: https://huggingface.co/datasets/LSW142857/OPSD-PI-SWE-Gym-1024.

sourceHugging Faceupdated 1mo agoView on Hugging Face
0likes18downloads
Dataset Card

OPSD-PI SWE-Gym 1024

OPSD-PI is a stage-adaptive privileged-information distillation project for repository-level coding agents. The student sees only the ordinary task and interaction history; a frozen teacher additionally receives stage-specific PI for EXPLORE, REPRODUCE, DIAGNOSE, EDIT, or VERIFY. Top-K forward KL distills that signal into the student LoRA and MTP head, while evaluation runs without external PI.

This repository publishes the prepared teacher-side inputs and portable source release for the Qwen3.5-9B OPSD-PI Weak experiment. The training implementation uses slime, Megatron-LM, SGLang, mini-swe-agent, and Apptainer as runtime components, but the maintained project and runbook are OPSD-PI-specific.

Contents

  • —data/train.jsonl: 1024 unique SWE-Gym tasks with stage-adaptive PI.
  • —data/manifest.json: source hashes, ordered instance IDs, and validation metadata.
  • —release/: portable source archive and fresh-run installer.

Checksums:

text
d33c069a715ff6d4cc115ec7ceee024b5d792f9ee7a3182c1d7cb7ddf56e086e  data/train.jsonl
529ed48d4d130249d156d497404bd90d9afe7afebb3af9bf915a7bfa3106ff3b  data/manifest.json

The source swegym_1024.zip has SHA-256 43fb5b98792df0440d36cc985a8f59547e4ac156188903d83654abf7d62f7184. It contains every task from the earlier 512-task swegym.zip plus 512 new tasks. The overlapping 512 PI payloads are identical. All 1024 artifacts passed the required structure and quality checks and matched the SWE-Gym baseline.

Run

Use the maintained branch and its beginner guide:

bash
git clone --branch opsd-pi-training --single-branch \
  https://github.com/LeiLiLab/privilege-code-opsd.git OPSD_pi
cd OPSD_pi

OPSD_DATASET_ROWS=1024 \
OPSD_PYTHON=/absolute/path/to/cuda-env/bin/python \
  bash scripts/opsd_pi_bootstrap_public.sh /persistent/opsd

Full instructions: <https://github.com/LeiLiLab/privilege-code-opsd/tree/opsd-pi-training>

This dataset contains privileged teacher information and benchmark patches. Keep it teacher-only. Do not inject it into the student prompt during SWE-bench evaluation.