nebius/Qwen3-235B-Instruct-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: generated_message
struct:
- name: content
dtype: string
- name: reasoning_content
dtype: 'null'
- name: role
dtype: string
- name: tool_calls
sequence: 'null'
- name: finish_reason
dtype: string
splits:
- name: train
num_bytes: 3666857766
num_examples: 659808
download_size: 2067010910
dataset_size: 3666857766
---

# Qwen3-235B-Instruct-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 Qwen3-235B-A22B-Instruct-2507 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 [Qwen/Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507) 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/Qwen3-235B-Instruct-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}
}
```

### 数据集元信息
- 许可证：CC BY 4.0
- 任务类别：文本生成
- 语言：英语
- 配置项：
- 配置名称：default
数据文件：
- 拆分方式：训练集
路径：data/train-*
- 数据集信息：
- 特征：
1. 字段名：conversation，列表类型，包含子字段：
- 字段名：content，数据类型：字符串
- 字段名：role，数据类型：字符串
2. 字段名：generated_message，结构体类型，包含子字段：
- 字段名：content，数据类型：字符串
- 字段名：reasoning_content，数据类型：空值
- 字段名：role，数据类型：字符串
- 字段名：tool_calls，序列类型，元素类型：空值
3. 字段名：finish_reason，数据类型：字符串
- 拆分信息：
- 拆分方式：训练集，字节大小：3666857766，样本数量：659808
- 下载大小：2067010910
- 数据集总大小：3666857766

# Qwen3-235B-Instruct-Infinity-Instruct-0625

## 数据集描述
本数据集隶属于用于推测解码（speculative decoding）研究的LK-Speculators集合，共包含66万条提示词-回复对，用于训练与Qwen3-235B-A22B-Instruct-2507搭配使用的草稿模型（作为目标模型）。本数据集通过将温度参数设为1，使用[Qwen/Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507)对[Infinity-Instruct-0625](https://huggingface.co/datasets/BAAI/Infinity-Instruct)中的提示词生成回复构建而成。如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 Losses: Direct Acceptance Rate Optimization for Speculative Decoding](https://arxiv.org/abs/2602.23881)。

## 数据集结构
- **格式**：Parquet
- **样本条数**：659,808

## 用法
python
from datasets import load_dataset

dataset = load_dataset("nebius/Qwen3-235B-Instruct-Infinity-Instruct-0625")

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

## 引用

@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}
}