marauder-actual
151 lines
3.7 KiB
Markdown
151 lines
3.7 KiB
Markdown
# Serpent — Python Specialist LoRA
|
||
|
||
Adapter codename: `serpent`
|
||
Agent: `build-python`
|
||
Base model: Qwen/Qwen3.5-27B
|
||
|
||
## Objective
|
||
|
||
Teach the model Python-idiomatic code generation aligned with the build-python
|
||
agent's system prompt. The adapter should internalize:
|
||
|
||
- Type hints on all function signatures (params + return)
|
||
- `pathlib.Path` over `os.path`
|
||
- `uv` for dependency management, `pyproject.toml` as source of truth
|
||
- Pydantic v2 patterns (`.model_dump()`, not `.dict()`)
|
||
- `pytest` + `pytest-asyncio` for testing
|
||
- `ruff` for lint + format
|
||
- ruff → pytest → mypy verification cycle
|
||
|
||
## Data Sources
|
||
|
||
### Session extraction (~100–200 examples)
|
||
|
||
Classify from opencode `build` sessions by:
|
||
- File paths: `.py`, `pyproject.toml`, `requirements.txt`, `setup.py`
|
||
- Bash commands: `python`, `pip`, `uv`, `pytest`, `ruff`, `mypy`
|
||
- Error patterns: `SyntaxError`, `TypeError`, `ImportError`, traceback format
|
||
- Framework detection: FastAPI, Django, Flask imports
|
||
|
||
### Git repo mining (~50–100 examples)
|
||
|
||
Target repos:
|
||
- `madcat-os/lora` — training scripts (this repo)
|
||
- Automation scripts in `~/.config/opencode/scripts/`
|
||
- Any Python projects in madcat-os org
|
||
|
||
Extract patterns:
|
||
- Data processing scripts (JSONL, datasets)
|
||
- ML training pipelines (unsloth, torch, datasets)
|
||
- CLI tools and automation
|
||
- FastAPI endpoint implementations
|
||
|
||
### Synthetic (~50–100 examples)
|
||
|
||
- async/await with `asyncio` patterns
|
||
- Pydantic v2 model definitions with validators
|
||
- `pathlib` file operations
|
||
- Type-narrowing with `TypeGuard`, `assert_never`
|
||
- Dataclass vs Pydantic decision points
|
||
- `uv` project setup and dependency resolution
|
||
|
||
## Example Format
|
||
|
||
```json
|
||
{
|
||
"messages": [
|
||
{
|
||
"role": "system",
|
||
"content": "<build-python agent system prompt>"
|
||
},
|
||
{
|
||
"role": "user",
|
||
"content": "Add a CLI argument to specify output format (json/csv) in the extraction script"
|
||
},
|
||
{
|
||
"role": "assistant",
|
||
"content": null,
|
||
"tool_calls": [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "read",
|
||
"arguments": {"filePath": "extract.py"}
|
||
}
|
||
}
|
||
]
|
||
},
|
||
{
|
||
"role": "tool",
|
||
"content": "<existing script>"
|
||
},
|
||
{
|
||
"role": "assistant",
|
||
"content": null,
|
||
"tool_calls": [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "edit",
|
||
"arguments": {
|
||
"filePath": "extract.py",
|
||
"oldString": "def main():",
|
||
"newString": "def main(output_format: Literal[\"json\", \"csv\"] = \"json\") -> None:"
|
||
}
|
||
}
|
||
}
|
||
]
|
||
},
|
||
{
|
||
"role": "tool",
|
||
"content": "Edit applied"
|
||
},
|
||
{
|
||
"role": "assistant",
|
||
"content": null,
|
||
"tool_calls": [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "bash",
|
||
"arguments": {"command": "ruff check extract.py && ruff format --check extract.py"}
|
||
}
|
||
}
|
||
]
|
||
},
|
||
{
|
||
"role": "tool",
|
||
"content": "All checks passed"
|
||
},
|
||
{
|
||
"role": "assistant",
|
||
"content": "Added `--format` CLI arg via `argparse`. Type-hinted with `Literal[\"json\", \"csv\"]`. Ruff clean."
|
||
}
|
||
]
|
||
}
|
||
```
|
||
|
||
## Evaluation Criteria
|
||
|
||
1. All functions have type hints (params and return)
|
||
2. Uses `pathlib.Path`, not `os.path`
|
||
3. `ruff check` and `ruff format --check` pass
|
||
4. `pytest` tests pass
|
||
5. Pydantic v2 patterns (no v1 `.dict()`, `.json()`)
|
||
6. No `requirements.txt` — uses `pyproject.toml` + `uv`
|
||
7. Tool call sequence: read → edit → lint → test
|
||
|
||
## Training Config Overrides
|
||
|
||
```python
|
||
MAX_SEQ = 8192
|
||
LR = 5e-5
|
||
```
|
||
|
||
## Estimated Size
|
||
|
||
- 200–400 examples
|
||
- ~1.5M tokens
|
||
- Training time: ~1.5 hrs on H100
|
||
- Adapter size: ~305 MB
|