README.md

---
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
    - name: thinking
      dtype: 'null'
  - name: status
    dtype: string
  splits:
  - name: train
    num_bytes: 20379033667
    num_examples: 1418401
    download_size: 20358035838
    dataset_size: 20379033667
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Perfect Blend - Qwen3-8B Regenerated

## Dataset Description

This dataset is a **regenerated** version of the [Open Perfect Blend](https://huggingface.co/datasets/mlabonne/open-perfectblend) (Perfect Blend) data. Each conversation was re-generated by **Qwen/Qwen3-8B** as the target model, so that the responses follow the target model's output distribution. It is intended for training **speculative decoding draft models** (e.g., DFlash, EAGLE) that align with Qwen3-8B, such as in [SpecForge](https://github.com/volcengine/SpecForge).

- **Target model:** Qwen/Qwen3-8B  
- **Source:** Perfect Blend (mlabonne/open-perfectblend)  
- **Splits:** `train` (1.42M rows)  
- **Approximate size:** ~20.4 GB (Parquet)

### Data Structure

| Column          | Type   | Description                                       |
|-----------------|--------|---------------------------------------------------|
| `id`            | int64  | Sample index                                      |
| `conversations` | list   | List of turns with `role`, `content`, `thinking`  |
| `status`        | string | e.g. `"success"` for successful regeneration      |

Each item in `conversations` has:
- `role`: `"user"` or `"assistant"`
- `content`: text of the message
- `thinking`: optional thinking content (e.g. for reasoning models)

---

## How to Load / Download

### Using the `datasets` library

```python
from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("jihwan1205/perfectblend-qwen3-8b-regen", split="train")

# With streaming to avoid loading everything into memory
dataset = load_dataset(
    "jihwan1205/perfectblend-qwen3-8b-regen",
    split="train",
    streaming=True
)
```

### Using in SpecForge (e.g. DFlash online training)

Use this dataset as the training data path:

```bash
--train-data-path jihwan1205/perfectblend-qwen3-8b-regen
```

---

## License

Same as the source dataset [Open Perfect Blend](https://huggingface.co/datasets/mlabonne/open-perfectblend). Please check the original dataset page for license details.

---

## Citation

If you use this dataset, please cite the original Perfect Blend and the model used for regeneration (Qwen3-8B).