macadeliccc/distilabel-neurology-preferences-2k

---
dataset_info:
features:
- name: input
dtype: string
- name: generation_model
sequence: string
- name: generation_prompt
list:
list:
- name: content
dtype: string
- name: role
dtype: string
- name: raw_generation_responses
sequence: string
- name: generations
sequence: string
- name: labelling_model
dtype: string
- name: labelling_prompt
list:
- name: content
dtype: string
- name: role
dtype: string
- name: raw_labelling_response
dtype: string
- name: rating
sequence: float64
- name: rationale
sequence: string
splits:
- name: train
num_bytes: 36980005
num_examples: 2000
download_size: 12336689
dataset_size: 36980005
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Dataset Card for "distilabel-neurology-preferences-2k"

## Preprocessing used on distilabel preferences

```python
from datasets import load_dataset

# Load your dataset
dataset_identifier = "macadeliccc/distilabel-neurology-dpo"
dataset = load_dataset(dataset_identifier, split='train')

def process_item(item):
# Step 1: Identify the highest-rated generation
ratings = item['rating']
highest_rating_index = ratings.index(max(ratings))

# Step 2: Select the corresponding prompt
selected_prompt_pair = item['generation_prompt'][highest_rating_index]
system_message = next((prompt['content'] for prompt in selected_prompt_pair if prompt['role'] == 'system'), "")
user_query = next((prompt['content'] for prompt in selected_prompt_pair if prompt['role'] == 'user'), "")

# Step 3: Construct the combined prompt
prompt = f"{system_message}\n\n{user_query}"

# Select the chosen and rejected responses based on ratings
chosen = item['generations'][highest_rating_index]
rejected = [resp for i, resp in enumerate(item['generations']) if i != highest_rating_index]

return {
"prompt": prompt,
"chosen": chosen,
"rejected": rejected
}

# Apply the processing function to each item in the dataset
transformed_dataset = dataset.map(process_item)

# Example of inspecting the first transformed item
print(transformed_dataset[0])
```

## Prompt format

For use during fine tuning

```python
from datasets import load_dataset

dataset_identifier = "macadeliccc/distilabel-neurology-dpo"
disti_neuro = load_dataset(dataset_identifier, split='train')

def chatml_format(disti_neuro):
# get everything except the last message as input
prompt = tokenizer.apply_chat_template(disti_neuro["prompt"][:-1], tokenize=False)
# get the last assistant responses
chosen = example["chosen"][-1]["content"] + "</s>"
rejected = example["rejected"][-1]["content"] + "</s>"

return {
"prompt": system + prompt,
"chosen": chosen,
"rejected": rejected,
}

# Save columns
original_columns = disti_neuro.column_names

# Format dataset
disti_neuro = disti_neuro.map(
chatml_format,
remove_columns=original_columns
)

```

原始信息汇总

数据集信息

特征

• input: 类型为字符串。
• generation_model: 序列类型为字符串。
• generation_prompt: 列表类型，包含以下子项：
  • content: 类型为字符串。
  • role: 类型为字符串。
• raw_generation_responses: 序列类型为字符串。
• generations: 序列类型为字符串。
• labelling_model: 类型为字符串。
• labelling_prompt: 列表类型，包含以下子项：
  • content: 类型为字符串。
  • role: 类型为字符串。
• raw_labelling_response: 类型为字符串。
• rating: 序列类型为浮点数。
• rationale: 序列类型为字符串。

分割

• train: 包含2000个样本，总字节数为36980005。

大小

• download_size: 12336689字节。
• dataset_size: 36980005字节。

配置

• default: 数据文件路径为data/train-*。