Domain Search MCP

# Domain Name QLoRA Training

This folder contains a minimal QLoRA fine-tuning setup for
`Qwen2.5-7B-Instruct` using the JSONL dataset generated by the MCP repo.

## Dataset
- Expected format: JSONL with `prompt` and `response` fields.
- Example line:
  ```json
  {"prompt":"Generate 10 brandable names for a crypto task platform...","response":"- name.com — reason\\n- name.ai — reason"}
  ```
- Default path: `data/domain-dataset-100k.jsonl`

## Quick start (GPU box)
```bash
python -m venv .venv
source .venv/bin/activate
pip install -r training/requirements.txt

cp training/.env.example training/.env
# Edit training/.env and set HF_TOKEN if required by the model host.

set -a
source training/.env
set +a

python training/qlora_train.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --data data/domain-dataset-100k.jsonl \
  --output training/output \
  --max_seq_len 512 \
  --batch_size 8 \
  --grad_accum 4 \
  --epochs 1
```

## Notes
- QLoRA uses 4-bit base weights + LoRA adapters.
- Output folder contains adapter weights and trainer logs.
- To change styles or constraints, regenerate the dataset first.