nanigock/repopeft-gru-commits-v2

Code2LoRA-GRU commit-level dataset, v2

This is the v2 snapshot of nanigock/repopeft-gru-commits extended with two precomputed embedding columns on every commit row:

• diff_embedding - 2048-d float16, encodes the filtered production_code_diff for that commit.
• repo_state_embedding - 2048-d float16, encodes the entire .py tree of the repository at that commit.

Both embeddings come from a single frozen model (Qwen/Qwen3-Embedding-0.6B) and are produced with the exact same recipe used by the published Code2LoRA-GRU and Code2LoRA-direct trainers. See EMBEDDINGS_README.json for the exact hyperparameters and per-split sha256 sums.

Why precompute?

The encoder is frozen for both Code2LoRA-GRU_commit and Code2LoRA_direct, so embedding cost is the dominant per-epoch overhead. Shipping them inside the dataset means the GRU trainer becomes a pure dataloader -> RNN -> LoRA-head pipeline (no Qwen3 forward pass on the hot path), and the static model can train without ever touching the encoder at all.

Memory-light loading recipe

The two embedding columns add ~8 GB total over the v1 dataset. For training rigs with limited host RAM, load them on demand::

import pyarrow.dataset as pads

ds = pads.dataset("commits/train.parquet", format="parquet")
scanner = ds.scanner(columns=["repo_id", "commit_index", "commit_sha",
                              "diff_embedding"],
                     batch_size=2048)
for batch in scanner.to_batches():
    ...

This streams one row group at a time (~250 MB peak) and never materializes the 1.4 GB diff column in memory.

Splits

(Identical to v1.) Repos are partitioned into a cross-repo train/cr_val/cr_test split; commits inside each repo are then chopped 80 / 10 / 10 chronologically into in-repo train / val / test slices.

Schemas

commits

column	type	description
repo_id	string	<owner>/<repo>
cross_repo_split	string	train / cr_val / cr_test
commit_index	int32	0-based index within the kept sequence
commit_sha	string	git SHA of this kept commit
commit_timestamp	string	ISO 8601
in_repo_split	string	train / val / test (80/10/10)
production_code_diff	large_string	filtered unified diff vs prev kept commit (test hunks removed)
n_new_assertions	int32	number of assertion events introduced
n_added_assertions	int32	events newly added at this commit
n_modified_assertions	int32	events modified at this commit
diff_embedding	list[float16, 2048]	Qwen3 encoding of production_code_diff. concat(MaxPool, MeanPool); not normalized.
repo_state_embedding	list[float16, 2048]	Qwen3 encoding of the full .py tree at this commit. concat(mean_files, max_files), L2-normalized.

qna

Unchanged from v1. See the v1 dataset card for the column list.

Reproducibility

The v2 build pipeline lives under create_dataset/ of the project repository:

• build_diff_embeddings_shard.py
• build_repo_state_embeddings_shard.py
• merge_gru_v2_embeddings.py

Per-shard SLURM launchers are in scripts/slurm/.

Citation

@misc{repopeft_gru_commits_v2_2026,
  title  = {Code2LoRA-GRU commit-level dataset, v2 (Qwen3 diff and repo-state embeddings)},
  year   = {2026},
  author = {RepoPeftData authors},
}