Felladrin/ChatML-truthy-dpo-v0.1

---
license: cc-by-4.0
language:
- en
size_categories:
- 1K<n<10K
---

[jondurbin/truthy-dpo-v0.1](https://huggingface.co/datasets/jondurbin/truthy-dpo-v0.1) in ChatML format, ready to use in [HuggingFace TRL's DPO Trainer](https://huggingface.co/docs/trl/main/en/dpo_trainer).

Python code used for conversion:

```python
from datasets import load_dataset

dataset = load_dataset("jondurbin/truthy-dpo-v0.1", split="train")

def format(columns):
prompt = f"<|im_start|>user\n{columns['prompt']}<|im_end|>\n<|im_start|>assistant\n"

if (columns['system']):
prompt = f"<|im_start|>system\n{columns['system']}<|im_end|>\n{prompt}"

return {
"prompt": prompt,
"chosen": f"{columns['chosen']}<|im_end|>",
"rejected": f"{columns['rejected']}<|im_end|>",
}

dataset.map(format).select_columns(['prompt', 'chosen', 'rejected', 'id', 'source']).to_parquet("train.parquet")
```

许可证：知识共享署名4.0（CC BY 4.0）
语言：英语
样本量区间：1000 < 样本数 < 10000

本数据集为[jondurbin/truthy-dpo-v0.1](https://huggingface.co/datasets/jondurbin/truthy-dpo-v0.1)的ChatML（Chat Markup Language）格式版本，可直接用于[HuggingFace TRL的DPO Trainer（Direct Preference Optimization Trainer）](https://huggingface.co/docs/trl/main/en/dpo_trainer)。

用于格式转换的Python代码如下：
python
from datasets import load_dataset

dataset = load_dataset("jondurbin/truthy-dpo-v0.1", split="train")

def format(columns):
prompt = f"<|im_start|>user
{columns['prompt']}<|im_end|>
<|im_start|>assistant
"

if (columns['system']):
prompt = f"<|im_start|>system
{columns['system']}<|im_end|>
{prompt}"

return {
"prompt": prompt,
"chosen": f"{columns['chosen']}<|im_end|>",
"rejected": f"{columns['rejected']}<|im_end|>",
}

dataset.map(format).select_columns(['prompt', 'chosen', 'rejected', 'id', 'source']).to_parquet("train.parquet")

原始信息汇总

数据集概述

基本信息

• 许可证: cc-by-4.0
• 语言: 英语
• 数据量: 1K<n<10K

数据格式

• 格式: ChatML
• 用途: 适用于HuggingFace TRLs DPO Trainer

数据转换代码

python from datasets import load_dataset

dataset = load_dataset("jondurbin/truthy-dpo-v0.1", split="train")

def format(columns): prompt = f"<|im_start|>user {columns[prompt]}<|im_end|> <|im_start|>assistant "

if (columns[system]):
    prompt = f"<|im_start|>system

{columns[system]}<|im_end|> {prompt}"

return {
    "prompt": prompt,
    "chosen": f"{columns[chosen]}<|im_end|>",
    "rejected": f"{columns[rejected]}<|im_end|>",
}

dataset.map(format).select_columns([prompt, chosen, rejected, id, source]).to_parquet("train.parquet")