RLHFlow/Prometheus2-preference-standard

---
dataset_info:
features:
- name: rejected
list:
- name: content
dtype: string
- name: role
dtype: string
- name: rejected_score
dtype: string
- name: chosen_score
dtype: string
- name: chosen
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 2396595452
num_examples: 199760
download_size: 356488395
dataset_size: 2396595452
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

Directly transformed from [prometheus-eval/Preference-Collection](https://huggingface.co/datasets/prometheus-eval/Preference-Collection).

Here is the data processing code:

```python
ds = load_dataset("prometheus-eval/Preference-Collection",split='train')
new_data = []
for example in tqdm(ds):
prompt = example['instruction']
responses = [example['orig_response_A'], example['orig_response_B']]
scores = [example['orig_score_A'], example['orig_score_B']]
chosen_idx = int(example['orig_preference'] == 'B') # A:0, B:1
chosen_response, rejected_response = responses[chosen_idx], responses[1-chosen_idx]
chosen_score, rejected_score = scores[chosen_idx], scores[1-chosen_idx]
chosen = [{"content": prompt, "role": "user"}, {"content": chosen_response, "role": "assistant"}]
rejected = [{"content": prompt, "role": "user"}, {"content": rejected_response, "role": "assistant"}]
row = {'rejected': rejected, 'rejected_score': rejected_score, 'chosen_score': chosen_score,'chosen': chosen,}
new_data.append(row)
new_ds = Dataset.from_list(new_data)
```

原始信息汇总

数据集概述

数据集特征

• rejected
  • content: 数据类型为字符串
  • role: 数据类型为字符串
• rejected_score: 数据类型为字符串
• chosen_score: 数据类型为字符串
• chosen
  • content: 数据类型为字符串
  • role: 数据类型为字符串

数据集分割

• train
  • num_bytes: 2396595452
  • num_examples: 199760

数据集大小

• download_size: 356488395
• dataset_size: 2396595452

配置

• config_name: default
  • data_files
    • split: train
      • path: data/train-*