v2 README
Browse files
README.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# SWE-Bench Trajectory Eval Bundle (v2)
|
| 6 |
+
|
| 7 |
+
Airgapped-ready bundle for the trajectory-probe downstream eval.
|
| 8 |
+
|
| 9 |
+
## Changes from v1
|
| 10 |
+
|
| 11 |
+
- **django dropped** (230 base_commits × 5 agents = 1139 trajectories removed).
|
| 12 |
+
django's blobs were too slow to prefetch at scale; the other 11 repos
|
| 13 |
+
give 1317 trajectories which is plenty.
|
| 14 |
+
- **Sparse-checkout blobs only**. For each (repo, base_commit), only the
|
| 15 |
+
files any trajectory's patch touches are fetched. Dramatically smaller
|
| 16 |
+
than the v1 "full-tree" approach; still enough for the graph encoder
|
| 17 |
+
to see the patched files + their containing package.
|
| 18 |
+
|
| 19 |
+
## Contents (570 MB compressed)
|
| 20 |
+
|
| 21 |
+
- `specs.jsonl`: 1317 SWE-Bench Verified agent trajectories (down from
|
| 22 |
+
2456). 46% resolved. 11 repos, 269 unique base_commits.
|
| 23 |
+
- `repos/`: 11 blobless clones with sparse-checkout already configured
|
| 24 |
+
per base_commit. Working trees on-disk reflect the UNION of touched
|
| 25 |
+
paths across all trajectories for each base_commit.
|
| 26 |
+
- `graphjepa/`: pipeline code (trajectory_pipeline with `extract_patch_paths`
|
| 27 |
+
and `set_sparse_paths`, trajectory_realize, trajectory_probe, checks).
|
| 28 |
+
- `harvest.log`.
|
| 29 |
+
|
| 30 |
+
## Dataset
|
| 31 |
+
|
| 32 |
+
| Repo | Trajs | Resolved |
|
| 33 |
+
|---|---|---|
|
| 34 |
+
| sympy/sympy | 370 | 48% |
|
| 35 |
+
| sphinx-doc/sphinx | 207 | 32% |
|
| 36 |
+
| matplotlib | 167 | 41% |
|
| 37 |
+
| scikit-learn | 158 | 70% |
|
| 38 |
+
| astropy | 110 | 42% |
|
| 39 |
+
| pydata/xarray | 108 | 51% |
|
| 40 |
+
| pytest-dev/pytest | 95 | 49% |
|
| 41 |
+
| pylint-dev/pylint | 49 | 22% |
|
| 42 |
+
| psf/requests | 38 | 61% |
|
| 43 |
+
| mwaskom/seaborn | 10 | 0% |
|
| 44 |
+
| pallets/flask | 5 | 80% |
|
| 45 |
+
| **total** | **1317** | **46%** |
|
| 46 |
+
|
| 47 |
+
## Downstream workflow (on airgapped big)
|
| 48 |
+
|
| 49 |
+
```bash
|
| 50 |
+
huggingface-cli download IDMedicine/code-graph-trajeval-v2 traj_full_bundle.tar.gz \
|
| 51 |
+
--local-dir outputs/traj_real
|
| 52 |
+
|
| 53 |
+
cd /opt/train/code-transformer/code-transformer
|
| 54 |
+
tar -xzf outputs/traj_real/traj_full_bundle.tar.gz
|
| 55 |
+
rsync -a traj_full/graphjepa/ graphjepa/
|
| 56 |
+
mkdir -p outputs/traj_real
|
| 57 |
+
cp traj_full/specs.jsonl outputs/traj_real/
|
| 58 |
+
mv traj_full/repos outputs/traj_real/repos
|
| 59 |
+
|
| 60 |
+
export BERT_MODEL_PATH=/opt/train/datasets/bert-base-uncased
|
| 61 |
+
GPUS=0,2,3,4,5,6,7 bash graphjepa/scripts/trajectory_full.sh
|
| 62 |
+
tail -f outputs/traj_real/logs/prebuild_shard*_gpu*.log
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
Because the clone already has the right blobs prefetched on laptop,
|
| 66 |
+
the airgapped big machine can checkout every base_commit without any
|
| 67 |
+
network access.
|