# Training Data

This folder documents the task-specific datasets used to fine-tune
ToneBridge-MiniCPM4.1-8B.

The files are JSONL files in chat prompt/completion format.

## Files

| File | Rows | Purpose |
|---|---:|---|
| `hsk12_english_order_train.jsonl` | 1,800 | Training split for HSK 1/2-style sentences influenced by English word order. |
| `hsk12_english_order_valid.jsonl` | 200 | Validation split for the HSK 1/2 word-order task. |
| `context_tone_hsk3_train.jsonl` | 4,500 | Training split for HSK 1-3 context and tone adaptation. |
| `context_tone_hsk3_valid.jsonl` | 500 | Validation split for the HSK 1-3 context/tone task. |

## Dataset Format

Each row follows this structure:

```json
{
  "prompt": [
    {
      "role": "system",
      "content": "你是中文语境校对助手。只输出更合适的句子，不要解释。"
    },
    {
      "role": "user",
      "content": "上下文：...\n原句：...\n任务：请根据上下文把原句改成更合适的中文。/no_think"
    }
  ],
  "completion": [
    {
      "role": "assistant",
      "content": "..."
    }
  ]
}
```

## Notes

- The data is synthetic and task-specific.
- It is intended for beginner Mandarin correction and tone adaptation.
- The dataset was generated to reduce exact duplicate source/correction pairs.
- The data should not be treated as a general Chinese benchmark.