IDMedicine/code-graph-trajeval-v2

SWE-Bench Trajectory Eval Bundle (v2)

Airgapped-ready bundle for the trajectory-probe downstream eval.

Changes from v1

• —django dropped (230 base_commits × 5 agents = 1139 trajectories removed). django's blobs were too slow to prefetch at scale; the other 11 repos give 1317 trajectories which is plenty.
• —Sparse-checkout blobs only. For each (repo, base_commit), only the files any trajectory's patch touches are fetched. Dramatically smaller than the v1 "full-tree" approach; still enough for the graph encoder to see the patched files + their containing package.

Contents (570 MB compressed)

• —specs.jsonl: 1317 SWE-Bench Verified agent trajectories (down from 2456). 46% resolved. 11 repos, 269 unique base_commits.
• —repos/: 11 blobless clones with sparse-checkout already configured per basecommit. Working trees on-disk reflect the UNION of touched paths across all trajectories for each basecommit.
• —graphjepa/: pipeline code (trajectorypipeline with `extractpatchpaths` and `setsparsepaths`, trajectoryrealize, trajectory_probe, checks).
• —harvest.log.

Dataset

Downstream workflow (on airgapped big)

huggingface-cli download IDMedicine/code-graph-trajeval-v2 traj_full_bundle.tar.gz \
    --local-dir outputs/traj_real

cd /opt/train/code-transformer/code-transformer
tar -xzf outputs/traj_real/traj_full_bundle.tar.gz
rsync -a traj_full/graphjepa/ graphjepa/
mkdir -p outputs/traj_real
cp traj_full/specs.jsonl outputs/traj_real/
mv traj_full/repos outputs/traj_real/repos

export BERT_MODEL_PATH=/opt/train/datasets/bert-base-uncased
GPUS=0,2,3,4,5,6,7 bash graphjepa/scripts/trajectory_full.sh
tail -f outputs/traj_real/logs/prebuild_shard*_gpu*.log

Because the clone already has the right blobs prefetched on laptop, the airgapped big machine can checkout every base_commit without any network access.