README.md

metadata

license: mit
tags:
  - code
  - assertion-completion
  - repository-level

code2lora-data-smartcap — smart-capped GRU training QnAs

Smart-capped, file/function-balanced per-commit QnA dataset used to train Code2LoRA-GRU in the paper. The capping reduces the heavy-tailed training distribution by:

1. Dropping trivial targets (len(target.strip()) < 4, bare ")", etc.).
2. Round-robin sampling within each commit across (test_file, test_function) groups so file/function diversity is preserved.
3. Capping each commit at max_per_commit = 8 and each (commit, test_file) at max_per_file = 4.

Only in_repo_split == "train" rows are capped. Val and test rows pass through verbatim, so evaluation suites remain comparable to the uncapped dataset.

Counts

File	Rows	Notes
qna/cr_train.parquet	1,187,359	215,129 capped train + 437,227 val + 535,003 test (val/test untouched)
qna/cr_val.parquet	641,848	full uncapped
qna/cr_test.parquet	476,455	full uncapped

Split semantics

Two independent split axes:

• cross_repo_split ∈ {train, cr_val, cr_test} -- which repositories the model has seen.
• in_repo_split ∈ {train, val, test} -- chronological 80/10/10 split of each repo's commits.

The four evaluation suites used in the paper map onto these axes as follows:

Paper suite	File	Selection
Train (loss)	qna/cr_train.parquet	in_repo_split == "train"
IR val	qna/cr_train.parquet	in_repo_split == "val"
IR test	qna/cr_train.parquet	in_repo_split == "test"
CR val	qna/cr_val.parquet	all rows
CR test	qna/cr_test.parquet	all rows

All evaluation suites use the full uncapped assertion set. Only the in-repo train loss path is smart-capped (see code2lora/code2lora-data-smartcap).

Companion datasets

• code2lora/code2lora-data-commits per-commit metadata + diff embeddings
• code2lora/code2lora-data-snapshots direct-projection (static) variant