---
license: mit
tags: [code, assertion-completion, repository-level]
configs:
- config_name: default
  data_files:
  - split: train
    path: qna/train.parquet
  - split: ir_val
    path: qna/ir_val.parquet
  - split: ir_test
    path: qna/ir_test.parquet
  - split: cr_val
    path: qna/cr_val.parquet
  - split: cr_test
    path: qna/cr_test.parquet
  - split: ood_test
    path: qna/ood_test.parquet
---

# code2lora-data-snapshots — static-snapshot RepoPeftBench

One row per (repository, anchor commit) for the **direct-projection
Code2LoRA** trainer. Each commit row carries a 2048-dim
Qwen3-Embedding repo-state vector; each QnA row is a single assertion
extracted from a held-out test file at that snapshot.

## Layout

```
commits/
  train.parquet      400 anchor commits (one per train repo)
  ir_val.parquet     in-repo-val snapshots from train repos
  ir_test.parquet    in-repo-test snapshots from train repos
  cr_val.parquet     held-out cr-val repos
  cr_test.parquet    held-out cr-test repos
  ood_test.parquet   post-cutoff OOD repos
qna/
  <same suite names>   matching assertion rows for each commit row
```

Splits are **named directly by paper suite** -- no extra filtering needed.

## Companion datasets

* `code2lora/code2lora-data-commits`   per-commit metadata + diff embeddings (for GRU)
* `code2lora/code2lora-data-smartcap`  smart-capped per-commit QnAs (GRU training loss)
* `code2lora/code2lora-data-ood`       full OOD commit-derived split