README.md

---
dataset_info:
  features:
  - name: task_id
    dtype: string
  - name: repo
    dtype: string
  - name: file_path
    dtype: string
  - name: function_name
    dtype: string
  - name: qualified_name
    dtype: string
  - name: function_type
    dtype: string
  - name: class_name
    dtype: string
  - name: prompt
    dtype: string
  - name: signature
    dtype: string
  - name: docstring
    dtype: string
  - name: canonical_solution
    dtype: string
  - name: full_function
    dtype: string
  - name: tests
    dtype: string
  - name: setup
    dtype: string
  - name: metadata
    dtype: string
  - name: validation
    dtype: string
  - name: original_task_id
    dtype: string
  - name: full_context
    dtype: string
  splits:
  - name: train
    num_bytes: 9279426
    num_examples: 55
  download_size: 1911300
  dataset_size: 9279426
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# CodeDP Repo-Patch Benchmark (CPT-friendly)

Repository-level code completion benchmark for evaluating continual pre-training (CPT) models. 55 tasks from 12 real-world repositories, each requiring the model to generate a function body given file-level context.

## Prompt Format

The `prompt` field contains the file context **up to** the target function's signature and docstring (truncated at the original `# TODO: Implement this function` marker). This format is directly usable by base/completion models — the model simply continues generating the function body.

The `full_context` field preserves the original full-file prompt (including code after the target function) for reference or fill-in-the-middle approaches.

### Base/Completion Models

```python
from datasets import load_dataset
ds = load_dataset("melihcatal/codedp-bench-repo-patch-cpt", split="train")

# prompt is ready for completion — ends with signature + docstring
prompt = ds[0]["prompt"]
# Model generates the function body from here
```

### Instruction Models

For chat/instruction models, wrap the prompt in a chat template:

```python
msg = f"Complete the implementation of `{ds[0]['function_name']}`. Return ONLY the function body.\n\n```python\n{ds[0]['prompt']}\n```"
```

## Fields

| Field | Description |
|-------|-------------|
| `prompt` | File context up to function signature + docstring (CPT-ready) |
| `full_context` | Full file with `# TODO` marker and downstream code |
| `canonical_solution` | Reference function body |
| `signature` | Function signature |
| `docstring` | Function docstring (empty for 22/55 tasks) |
| `function_name` | Target function name |
| `class_name` | Enclosing class (if method, else null) |
| `tests` | JSON list of pytest test cases |
| `setup` | JSON with repo URL, install command, commit SHA |
| `full_function` | Complete function (signature + docstring + body) |
| `metadata` | JSON with body_lines, file_lines, has_docstring, num_tests |
| `validation` | Test validation status |

## Statistics

- **55 tasks** from 12 repositories
- **13 class methods**, 42 standalone functions
- **33 with docstrings**, 22 without
- Prompt lengths (after truncation): median ~2,800 chars (vs ~6,300 before)
- Reference body lengths: median 437 chars

## Metrics

Reference-based metrics (no repo setup needed):
- **BLEU-4**: Token-level BLEU score
- **CodeBLEU**: Syntax-aware code similarity
- **Edit Similarity**: 1 - normalized Levenshtein distance
- **Exact Match**: Normalized whitespace comparison

## Evaluation

```bash
python -m evaluation.utility.run_repo_patch \
    --model_path ./output/model/checkpoint-final \
    --benchmark_path melihcatal/codedp-bench-repo-patch-cpt \
    --output_dir results/repo_patch/model/variant \
    --devices auto --batch_size 4

# For instruction models, add --chat_template
```