nebius/gpt-oss-20b-Infinity-Instruct-0625

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: conversation
list:
- name: content
dtype: string
- name: role
dtype: string
- name: finish_reason
dtype: string
- name: generated_message
struct:
- name: annotations
dtype: 'null'
- name: audio
dtype: 'null'
- name: content
dtype: string
- name: function_call
dtype: 'null'
- name: reasoning_content
dtype: string
- name: refusal
dtype: 'null'
- name: role
dtype: string
- name: tool_calls
sequence: 'null'
splits:
- name: train
num_bytes: 6009080088
num_examples: 659808
download_size: 3493807994
dataset_size: 6009080088
---

# gpt-oss-20b-Infinity-Instruct-0625

## Dataset Description

This dataset is part of the LK-Speculators collection for speculative decoding research. It contains 660K prompt-response pairs designed for training draft models that are used alongside gpt-oss-20b as the target model. The dataset was created by generating responses to the prompts from [Infinity-Instruct-0625](https://huggingface.co/datasets/BAAI/Infinity-Instruct) with [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) at temperature=1.
For more details on the training methodology and results, see our paper: [LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding](https://arxiv.org/abs/2602.23881).

## Dataset Structure

- **Format**: parquet
- **Rows**: 659,808

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("nebius/gpt-oss-20b-Infinity-Instruct-0625")
```

## License

The dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

## Citation

```
@misc{samarin2026lklosses,
title = {LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding},
author = {Alexander Samarin and Sergei Krutikov and Anton Shevtsov and Sergei Skvortsov and Filipp Fisin and Alexander Golubev},
year = {2026},
eprint = {2602.23881},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2602.23881}
}
```

---
许可证: CC BY 4.0
任务类别:
- 文本生成（text-generation）
语言:
- 英语（en）
配置项:
- 配置名称: default
数据文件:
- 数据拆分: train
路径: data/train-*
数据集信息:
特征:
- 名称: conversation
类型: 列表
子项:
- 名称: content
数据类型: 字符串
- 名称: role
数据类型: 字符串
- 名称: finish_reason
数据类型: 字符串
- 名称: generated_message
类型: 结构体
子项:
- 名称: annotations
数据类型: 空值（null）
- 名称: audio
数据类型: 空值（null）
- 名称: content
数据类型: 字符串
- 名称: function_call
数据类型: 空值（null）
- 名称: reasoning_content
数据类型: 字符串
- 名称: refusal
数据类型: 空值（null）
- 名称: role
数据类型: 字符串
- 名称: tool_calls
类型: 序列，元素数据类型: 空值（null）
数据拆分:
- 名称: train
字节大小: 6009080088
样本数量: 659808
下载大小: 3493807994
数据集总大小: 6009080088
---

# gpt-oss-20b-Infinity-Instruct-0625

## 数据集描述
本数据集隶属于用于**推测式解码（speculative decoding）**研究的LK-Speculators数据集集合。其包含66万条提示词-回复对，用于训练配合gpt-oss-20b作为目标模型的草稿模型。本数据集通过以温度参数=1的设置，使用[openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) 为[Infinity-Instruct-0625](https://huggingface.co/datasets/BAAI/Infinity-Instruct)中的提示词生成回复而构建。若需了解训练方法与实验结果的更多细节，请参阅我们的论文：[LK损失：面向推测式解码的直接接受率优化](https://arxiv.org/abs/2602.23881)。

## 数据集结构
- **数据格式**：Parquet
- **样本总数**：659808

## 使用方法
python
from datasets import load_dataset

dataset = load_dataset("nebius/gpt-oss-20b-Infinity-Instruct-0625")

## 许可证
本数据集遵循[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)协议发布。

## 引用格式

@misc{samarin2026lklosses,
title = {LK损失：面向推测式解码的直接接受率优化},
author = {Alexander Samarin and Sergei Krutikov and Anton Shevtsov and Sergei Skvortsov and Filipp Fisin and Alexander Golubev},
year = {2026},
eprint = {2602.23881},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2602.23881}
}