RLHFlow/CodeUltraFeedback-standard

---
dataset_info:
features:
- name: rejected
list:
- name: content
dtype: string
- name: role
dtype: string
- name: chosen
list:
- name: content
dtype: string
- name: role
dtype: string
- name: chosen_score
dtype: string
- name: rejected_score
dtype: string
splits:
- name: train
num_bytes: 245682332
num_examples: 50156
download_size: 47661064
dataset_size: 245682332
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

The original dataset is coseal/CodeUltraFeedback. We provide the processing script as follows. In summarization, for each prompt, we include all possible comparisons, except those with the same rating.

```python
ds = load_dataset("coseal/CodeUltraFeedback", split="train")
import itertools
data = []
for example in ds:
prompt = example['instruction']
responses = {}
for row in example['responses']:
model_name, response = row['model'], row['response']
responses[model_name] = response
annotations = {}
for row in example['annotations']:
model_name, rating = row['model'], row['rating']
annotations[model_name] = rating
preference = 'code-' + example['preference']
all_models = []

for model_name in responses.keys():
if rating == 'N/A':
continue
all_models.append(model_name)

all_combinations = list(itertools.combinations(all_models, 2))

assert len(all_combinations) == len(all_models) * (len(all_models) - 1) / 2

for combination in all_combinations:
response1 = responses[combination[0]]
rating1 = annotations[combination[0]]
response2 = responses[combination[1]]
rating2 = annotations[combination[1]]

if rating1 == rating2:
continue

if rating1 > rating2:
chosen_message = [
{"content": prompt, "role": "user"},
{"content": response1, "role": "assistant"},
]
rejected_message = [
{"content": prompt, "role": "user"},
{"content": response2, "role": "assistant"},
]
chosen_rating = rating1
rejected_rating = rating2

elif rating1 < rating2:
chosen_message = [
{"content": prompt, "role": "user"},
{"content": response2, "role": "assistant"},
]
rejected_message = [
{"content": prompt, "role": "user"},
{"content": response1, "role": "assistant"},
]
chosen_rating = rating2
rejected_rating = rating1
else:
print("error")

data.append({"rejected": rejected_message, "chosen": chosen_message, "rejected_score": rejected_rating, "chosen_score": chosen_rating})

```

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

数据集分割

• train
  • num_bytes: 245682332
  • num_examples: 50156

数据集大小

• download_size: 47661064
• dataset_size: 245682332

配置

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