nanigock/repopeft-gru-commits

Code2LoRA-GRU commit-level dataset (smart-cap)

Per-repository commit sequences with test-file assertion QnA pairs, used to train the Code2LoRA-GRU_commit hypernetwork. The model ingests the filtered production-code diff at each commit, evolves a GRU hidden state across the commit history, and at each step generates a LoRA that adapts a base LLM (Qwen/Qwen2.5-Coder-1.5B) to the repository's current state. Targets are RHSs of newly-introduced assertions at each commit.

Splits

• Cross-repo split (cross_repo_split ∈ {train, cr_val, cr_test}): Each repository is assigned to exactly one of these sets. Held-out repositories are never seen during training.
• In-repo split (in_repo_split ∈ {train, val, test}): For each repo the kept commits are chopped 80 / 10 / 10 chronologically. Only the train slice contributes to the LoRA loss during training; the GRU is still allowed to ingest every diff in order.

Smart-cap filtering

The qna/train.parquet shipped here is the smart-cap snapshot of the raw QnA table. It applies three additive filters on the cross-repo training set:

1. Drop trivial targets (single-token bools / Nones / empty strings).
2. Per-(test_file, test_function) round-robin: at most max_per_function = 4 assertions per function.
3. Per-commit hard cap: at most max_per_commit = 8 distinct (file, function, target) triples per commit. The qna/cr_val.parquet and qna/cr_test.parquet files are unfiltered (verbatim copies of the canonical RepoPeftBench QnAs) so that cross-repo evaluation numbers stay comparable across model versions.

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 (chronological 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

qna

column	type	description
repo_id	string	<owner>/<repo>
commit_sha	string	commit at which this assertion event lives
commit_index	int32	matches the commits row
in_repo_split	string	train / val / test
cross_repo_split	string	train / cr_val / cr_test
test_file	string	path inside the repo, posix style
test_function	string	enclosing pytest / unittest function name
prefix	large_string	structured context: imports + class + helpers + test body up to the assertion
target	large_string	the assertion RHS to predict
assertion_type	string	pytest / unittest / numpy / etc.
assertion_event	string	added or modified
difficulty	string	easy / medium / hard (heuristic)

Usage

from datasets import load_dataset

commits = load_dataset("nanigock/repopeft-gru-commits", "commits")
qna     = load_dataset("nanigock/repopeft-gru-commits", "qna")

print(commits)
# DatasetDict({train, cr_val, cr_test})
print(qna["train"][0]["prefix"][:200])

Training code

The model and training script live at github.com/lilianahotsko/RepoPeftData. See scripts/slurm/train_code2lora_gru_commits.sh for the SLURM launcher and hypernetwork/train_code2lora_gru_commits.py for the trainer.

Citation

@misc{repopeft_gru_commits_2026,
  title  = {Code2LoRA-GRU: A commit-sequential hypernetwork for repository-aware LoRA adaptation},
  year   = {2026},
  author = {RepoPeftData authors},
}