tasksource/oasst2_pairwise_rlhf_reward

---
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
dataset_info:
features:
- name: lang
dtype: string
- name: parent_id
dtype: string
- name: prompt
dtype: string
- name: chosen
dtype: string
- name: rejected
dtype: string
- name: rank
sequence: float64
splits:
- name: train
num_bytes: 68638275
num_examples: 26971
- name: validation
num_bytes: 3355134
num_examples: 1408
download_size: 0
dataset_size: 71993409
---
# Dataset Card for "oasst2_pairwise_rlhf_reward"

```python
import pandas as pd
from datasets import load_dataset,concatenate_datasets, Dataset, DatasetDict
import numpy as np

dataset = load_dataset("OpenAssistant/oasst2")

df=concatenate_datasets(list(dataset.values())).to_pandas()
m2t=df.set_index("message_id")['text'].to_dict()
m2r=df.set_index("message_id")['role'].to_dict()
m2p=df.set_index('message_id')['parent_id'].to_dict()

m2history=dict() # message id to unrolled history
for k,v in m2p.items():
history=[k]
while history[-1] in m2p:
history+=[m2p[history[-1]]]
m2history[k]="\n".join([f"{m2r[m]}: {m2t[m]}" for m in history[::-1] if m])

d=dict()
for split in "train","validation":
df=dataset[split].to_pandas()
df['prompt']=df.parent_id.map(lambda x: m2history.get(x,''))
df=df[~df['rank'].isna()]

def agg(x):
x=list(x)
return [x[0],x[-1]]
df=df.groupby(['prompt',"parent_id",'lang'])[['text','rank']].agg(agg).reset_index()
df=df[df['rank'].map(lambda x:len(set(x))>1)]
df['chosen'] = df.apply(lambda x:x['text'][np.argmin(x['rank'])],axis=1)
df['rejected'] = df.apply(lambda x:x['text'][np.argmax(x['rank'])],axis=1)
d[split]=Dataset.from_pandas(df[['lang','parent_id','prompt','chosen','rejected']],preserve_index=False)

DatasetDict(d).push_to_hub('tasksource/oasst2_pairwise_rlhf_reward')
```

原始信息汇总

数据集概述

数据集名称

• 名称: oasst2_pairwise_rlhf_reward

数据集配置

• 默认配置:
  • 训练集路径: data/train-*
  • 验证集路径: data/validation-*

数据集特征

• 特征列表:
  • lang: 数据类型为字符串
  • parent_id: 数据类型为字符串
  • prompt: 数据类型为字符串
  • chosen: 数据类型为字符串
  • rejected: 数据类型为字符串
  • rank: 数据类型为浮点数序列

数据集分割

• 训练集:
  • 字节数: 68638275
  • 样本数: 26971
• 验证集:
  • 字节数: 3355134
  • 样本数: 1408

数据集大小

• 下载大小: 0
• 数据集大小: 71993409