andysalerno/rainbowfish-v1

---
dataset_info:
features:
- name: source
dtype: string
- name: input
dtype: string
- name: output
dtype: string
splits:
- name: train
num_bytes: 148983951
num_examples: 69980
download_size: 70573434
dataset_size: 148983951
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

## Formatting

Formatting is compliant with ChatML. "input" is the context and "output" is the expected model output to train on.

## Details

See the repo file `generate_dataset.py` for exactly this dataset was generated.

An opinionated and filtered mix of the following datasets:

- argilla/ultrafeedback-binarized-preferences-cleaned
- heegyu/glaive-function-calling-v2-formatted
- berkeley-nest/Nectar
- argilla/distilabel-math-preference-dpo

## argilla/ultrafeedback-binarized-preferences-cleaned

**Filter**:
- `chosen-rating == 5` AND
- `len(chosen) == 2`

## heegyu/glaive-function-calling-v2-formatted

**Filter**:
- `function_description != ''`

**Transforms**:
- Added a system message randomly selected from a pool of generic system messages.

## berkeley-nest/Nectar

**Filter**:
- has an answer with `rank == 1` AND
- `turns > 1` AND
- `good_natured == True` AND
- answer.to_lower() does not start with "i'm sorry"

## argilla/distilabel-math-preference-dpo

**Filter**:
- `chosen-rating >= 9`

**Transforms**:
- Added a system message randomly selected from a pool of generic system messages.
- Removed the phrase "Take a deep breath, think step by step, and give an accurate response"

## Global formatting

All the above datasets were formatted to comply with ChatML.

原始信息汇总

数据集信息

特征

• source: 数据来源，类型为字符串。
• input: 输入内容，类型为字符串。
• output: 输出内容，类型为字符串。

数据分割

• train: 训练集，包含148,983,951字节，69,980个样本。

数据大小

• 下载大小: 70,573,434字节
• 数据集大小: 148,983,951字节

配置

• default: 默认配置，包含训练集数据文件路径为data/train-*。

数据集生成细节

数据集来源

• argilla/ultrafeedback-binarized-preferences-cleaned:

  • 过滤条件: chosen-rating == 5 且 len(chosen) == 2

• heegyu/glaive-function-calling-v2-formatted:

  • 过滤条件: function_description !=
  • 转换操作: 随机添加一个系统消息。

• berkeley-nest/Nectar:

  • 过滤条件: 包含一个rank == 1的答案，且turns > 1，good_natured == True，答案不以"Im sorry"开头。

• argilla/distilabel-math-preference-dpo:

  • 过滤条件: chosen-rating >= 9
  • 转换操作: 随机添加一个系统消息，移除短语"Take a deep breath, think step by step, and give an accurate response"。

全局格式

所有上述数据集都已格式化为符合ChatML标准。