README.md

---
language:
  - en
license: other
task_categories:
  - text-generation
tags:
  - python
  - code
  - pretraining
  - qwen3.5
  - conversation
  - instruction-tuning
size_categories:
  - 10K
---

# CodeParrot Clean Python (sample)

## Description

Teaches domain-specific instruction following and code generation for this expert.

## Source

- codeparrot/codeparrot-clean - 500 MB streamed sample of 12.8 GB; the per-file licence field is preserved in the text

Formatted for the MoE-orchestrator project
(https://github.com/michaelowusuntim6/MoE-orchestrator). Expert target:
`code_python`.

## Format

Each record is a JSON object with a `messages` field formatted for Qwen3.5's
native chat template:

```json
{"messages": [
  {"role": "system", "content": "..."},
  {"role": "user", "content": "..."},
  {"role": "assistant", "content": "..."}
]}
```

The records are consumed via `tokenizer.apply_chat_template()`. Special tokens
(`<|im_start|>`, `<|im_end|>`) are added by the template, never embedded in
content.

## Splits

- `train`: 47,048 records
- `val`: 975 records

## Usage

```python
from datasets import load_dataset
ds = load_dataset("michaelowusuntim6/python-codeparrot-qwen35", split="train")
print(ds[0]["messages"])
```

## License

per-file (mostly MIT/Apache-2.0). Upstream sources keep their own licences - see the source list
above and `docs/DATASET_SOURCES.md` in the MoE-orchestrator repository for
per-source detail.