hqq_plus_plus_mix_40k

# Dataset Card for Dataset Name

Replication of HQQ++ dataset mixture mentioned in [here](https://mobiusml.github.io/1bit_blog/#datasets) with Llama 3 Instruct chat template.

## Uses

Quantization Aware Training.

### Source Data

- [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco)
- [microsoft/orca-math-word-problems-200k](https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k)
- [meta-math/MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA)
- [HuggingFaceH4/ultrafeedback_binarized](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized)

#### Data Collection and Processing

```python
train_ds1 = load_dataset("timdettmers/openassistant-guanaco")['train']

train_ds2 = load_dataset("microsoft/orca-math-word-problems-200k")
train_ds2 = train_ds2['train'].shuffle(42).select(range(10000))

train_ds3 = load_dataset("meta-math/MetaMathQA")
train_ds3 = train_ds3['train'].shuffle(42).select(range(10000))

train_ds4 = load_dataset('HuggingFaceH4/ultrafeedback_binarized')
train_ds4 = train_ds4['train_sft'].shuffle(42).select(range(10000))
```

## Dataset Card Contact

Kerem Turgutlu: [e-mail]

# 数据集卡片：[数据集名称]
复现[此处](https://mobiusml.github.io/1bit_blog/#datasets)提及的HQQ++混合数据集，并适配Llama 3 Instruct对话模板。

## 应用场景
量化感知训练（Quantization Aware Training）。

### 源数据集
- [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco)
- [microsoft/orca-math-word-problems-200k](https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k)
- [meta-math/MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA)
- [HuggingFaceH4/ultrafeedback_binarized](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized)

#### 数据采集与处理流程
python
train_ds1 = load_dataset("timdettmers/openassistant-guanaco")['train']

train_ds2 = load_dataset("microsoft/orca-math-word-problems-200k")
train_ds2 = train_ds2['train'].shuffle(42).select(range(10000))

train_ds3 = load_dataset("meta-math/MetaMathQA")
train_ds3 = train_ds3['train'].shuffle(42).select(range(10000))

train_ds4 = load_dataset('HuggingFaceH4/ultrafeedback_binarized')
train_ds4 = train_ds4['train_sft'].shuffle(42).select(range(10000))

## 数据集卡片联系方式
凯雷姆·图尔古特卢（Kerem Turgutlu）：[e-mail]