nanigock/repopeft-code2lora-snapshots

Code2LoRA snapshots dataset

This dataset is the static-hypernetwork companion to nanigock/repopeft-gru-commits-v2. Every example is a single (repo, commit) snapshot annotated with:

• a 2048-d repo_state_embedding (frozen Qwen/Qwen3-Embedding-0.6B, concat of mean and max pooled file vectors, L2-normalized; details below);
• a list of canonical QnA pairs (test-file assertions) live at that commit.

It is designed for the direct Code2LoRA baseline: a single feed-forward hypernetwork maps repo_state_embedding -> LoRA delta, with no GRU rollout and no diff handling. Per-commit evaluation gives a decay curve directly comparable to the GRU.

Splits

split	rows (commits)	role
train	400	one anchor commit per train repo (last in_repo_split=='train').
ir_val	~3 k	train repos, in_repo_split=='val' commits.
ir_test	~6 k	train repos, in_repo_split=='test' commits.
cr_val	~9 k	held-out cr_val repos, all kept commits.
cr_test	~7 k	held-out cr_test repos, all kept commits.

The train QnAs are re-extracted at the anchor commit with the v2 extractor: extract_from_file -> select_balanced_pairs (max_per_repo=200, max_per_function=5, max_per_file=20). This guarantees that every QnA is actually present in the repo at the snapshot point (no stale assertions). The eval QnAs are the canonical RepoPeftBench QnAs, identical to those used to score Code2LoRA-GRU_commit.

repo_state_embedding details

hyperparameter	value
Model	Qwen/Qwen3-Embedding-0.6B
File chunking	2048 tokens, 256 overlap
Min window tokens	8
Per-chunk pooling	attention-mean of last_hidden_state
Per-file pooling	mean over chunk vectors (1024-d)
Per-repo pooling	concat(mean_files, max_files) (2048-d)
Repo vector norm	L2 normalized
File filter	tracked .py blobs, size <= 2 MB
Identical files dedup	by git ls-tree blob SHA

Loading

from datasets import load_dataset
commits = load_dataset("nanigock/repopeft-code2lora-snapshots", "commits")
qna     = load_dataset("nanigock/repopeft-code2lora-snapshots", "qna")

Join the two on (repo_id, commit_sha) to form (repo_state_embedding, prefix, target) triples for static training.

Citation

@misc{repopeft_code2lora_snapshots_2026,
  title  = {Code2LoRA snapshots: a static-hypernetwork dataset for repository-aware LoRA generation},
  year   = {2026},
  author = {RepoPeftData authors},
}