leideng/Dolci-Think-RL-7B-4K-Plus

---
dataset_info:
features:
- name: ground_truth
list: string
- name: dataset
list: string
- name: custom_id
dtype: string
- name: original_dataset
dtype: string
- name: outputs
list: string
- name: total_rollouts
dtype: int64
- name: total_correct_rollouts
dtype: float64
- name: passrate
dtype: float64
- name: dataset_source
dtype: string
- name: input_ids_prompt
list: int64
- name: input_ids
list: int32
- name: attention_mask
list: int8
- name: labels
list: int64
- name: prompt
dtype: string
- name: id
dtype: string
- name: key
dtype: string
- name: constraint_type
dtype: string
- name: constraint
dtype: string
- name: conversation_hash
dtype: string
- name: model
dtype: string
- name: predicted_label
dtype: string
splits:
- name: train
num_bytes: 4083445248
num_examples: 102014
download_size: 1893783057
dataset_size: 4083445248
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# Dolci-Think-RL-7B

## Dataset Summary

**Dolci-Think-RL-7B** is the reinforcement learning dataset used to train the *Olmo-3-7B-Think* model.
It contains **102,014** prompts designed to elicit deep reasoning across:

- Math
- Coding
- Precise Instruction Following
- General Chat

It blends high-quality curated sources with filtering designed for deliberate reasoning.

---

## Dataset Composition

### **Total Samples:** 102,014

### **Original Dataset Contribution**
| Source Dataset | Count |
|----------------|-------|
| IF Multi-Constraint | 29,813 |
| OMEGA Math ([paper](https://arxiv.org/abs/2506.18880)) | 15,000 |
| AceCoder ([paper](https://arxiv.org/abs/2502.01718)) | 10,107 |
| Tulu 3 Rewritten ([paper](https://arxiv.org/abs/2411.15124)) | 7,109 |
| Multi-Subject RLVR ([paper](https://arxiv.org/abs/2503.23829v1)) | 7,106 |
| AceReason-Math ([paper](https://arxiv.org/abs/2505.16400)) | 6,598 |
| WildChat English ([paper](https://arxiv.org/abs/2405.01470)) | 6,421 |
| KlearReasoner Code | 6,272 |
| SYNTHETIC-2 / PrimeIntellect ([blog](https://www.primeintellect.ai/blog/synthetic-2)) | 3,000 |
| MathSub-30K (KlearReasoner Math) ([paper](https://arxiv.org/abs/2508.07629)) | 2,999 |
| ORZ Math ([paper](https://arxiv.org/abs/2503.24290)) | 2,999 |
| DAPO-Math ([paper](https://arxiv.org/abs/2503.14476)) | 2,584 |
| Llama-Nemotron Post-Training Dataset ([paper](https://arxiv.org/abs/2505.00949)) | 2,006 |

### **Dataset Source Counts (Grouped Mixes)**
| Mix | Count |
|------|-------|
| Math RLVR Mixture | 30,180 |
| IF RLVR Mixture | 29,813 |
| Code RLVR Mixture | 21,385 |
| General RLVR Mixture | 20,636 |

---

## Data Sources & Description

### **Instruction Following**
- Up to 5 constraints
- Derived from IFBench-Train & IFEval-style tasks
- Filtered for clarity and non-toxicity

### **Math Reasoning**
- **OMEGA**
- **AceReason-Math**
- **ORZ Math**
- **DAPO-Math**
- **MathSub-30K**
- Wide domain coverage: geometry, algebra, combinatorics, proofs, etc.

### **Code Reasoning**
Includes four major families:
- **AceCoder**
- **KlearReasoner-Code**
- **SYNTHETIC-2 / PrimeIntellect**
- **Llama-Nemotron Post-Training Dataset**
All filtered via test-case execution.

### **General Long-Form Reasoning**
- Multi-Subject RLVR
- Tulu 3 rewritten (filtered via F1-score)
- WildChat English (filtered for reasoning suitability)

---

## Processing & Filtering

- **Execution-based code filtering** (test-case validated)
- **Topic filtering** for safety and quality
- **F1-based rewrite filtering** (Tulu 3)
- **Difficulty-tiered Nemotron subsets**
- **Strict deduplication**
- **Constraint normalization**

---

## License
This dataset is licensed under ODC-BY. It is intended for research and educational use in accordance with [Ai2's Responsible Use Guidelines](https://allenai.org/responsible-use).

## Citation

```
@misc{olmo2025olmo3,
title={Olmo 3},
author={Team Olmo and Allyson Ettinger and Amanda Bertsch and Bailey Kuehl and David Graham and David Heineman and Dirk Groeneveld and Faeze Brahman and Finbarr Timbers and Hamish Ivison and Jacob Morrison and Jake Poznanski and Kyle Lo and Luca Soldaini and Matt Jordan and Mayee Chen and Michael Noukhovitch and Nathan Lambert and Pete Walsh and Pradeep Dasigi and Robert Berry and Saumya Malik and Saurabh Shah and Scott Geng and Shane Arora and Shashank Gupta and Taira Anderson and Teng Xiao and Tyler Murray and Tyler Romero and Victoria Graf and Akari Asai and Akshita Bhagia and Alexander Wettig and Alisa Liu and Aman Rangapur and Chloe Anastasiades and Costa Huang and Dustin Schwenk and Harsh Trivedi and Ian Magnusson and Jaron Lochner and Jiacheng Liu and Lester James V. Miranda and Maarten Sap and Malia Morgan and Michael Schmitz and Michal Guerquin and Michael Wilson and Regan Huff and Ronan Le Bras and Rui Xin and Rulin Shao and Sam Skjonsberg and Shannon Zejiang Shen and Shuyue Stella Li and Tucker Wilde and Valentina Pyatkin and Will Merrill and Yapei Chang and Yuling Gu and Zhiyuan Zeng and Ashish Sabharwal and Luke Zettlemoyer and Pang Wei Koh and Ali Farhadi and Noah A. Smith and Hannaneh Hajishirzi},
year={2025},
eprint={2512.13961},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2512.13961},
}
```

dataset_info:
features:
- name: 真实标签（ground_truth）
list: 字符串列表（string）
- name: 数据集（dataset）
list: 字符串列表（string）
- name: 自定义ID（custom_id）
dtype: 字符串（string）
- name: 原始数据集（original_dataset）
dtype: 字符串（string）
- name: 模型输出（outputs）
list: 字符串列表（string）
- name: 总推演次数（total_rollouts）
dtype: 64位整数（int64）
- name: 正确推演次数（total_correct_rollouts）
dtype: 64位浮点数（float64）
- name: 通过率（passrate）
dtype: 64位浮点数（float64）
- name: 数据集来源（dataset_source）
dtype: 字符串（string）
- name: 提示词输入Token序列（input_ids_prompt）
list: 64位整数列表（int64）
- name: 输入Token序列（input_ids）
list: 32位整数列表（int32）
- name: 注意力掩码（attention_mask）
list: 8位整数列表（int8）
- name: 标签（labels）
list: 64位整数列表（int64）
- name: 提示词（prompt）
dtype: 字符串（string）
- name: 样本ID（id）
dtype: 字符串（string）
- name: 键值（key）
dtype: 字符串（string）
- name: 约束类型（constraint_type）
dtype: 字符串（string）
- name: 约束条件（constraint）
dtype: 字符串（string）
- name: 对话哈希值（conversation_hash）
dtype: 字符串（string）
- name: 模型（model）
dtype: 字符串（string）
- name: 预测标签（predicted_label）
dtype: 字符串（string）
splits:
- name: 训练集（train）
num_bytes: 4083445248
num_examples: 102014
download_size: 1893783057
dataset_size: 4083445248
configs:
- config_name: 默认配置（default）
data_files:
- split: 训练集（train）
path: data/train-*

# Dolci-Think-RL-7B

## 数据集概述

**Dolci-Think-RL-7B** 是用于训练*Olmo-3-7B-Think*模型的强化学习数据集。
其包含**102,014**条提示词，用于触发涵盖以下四类的深度推理任务：
- 数学推理
- 代码编程
- 精准指令遵循
- 通用对话

该数据集融合了高质量精选数据源，并针对严谨推理需求进行了针对性筛选。

---

## 数据集构成

### 总样本量：102,014

### 原始数据集贡献
| 源数据集 | 样本数 |
|----------------|-------|
| IF多约束（IF Multi-Constraint） | 29,813 |
| OMEGA数学数据集（OMEGA Math，[论文](https://arxiv.org/abs/2506.18880)） | 15,000 |
| AceCoder数据集（AceCoder，[论文](https://arxiv.org/abs/2502.01718)） | 10,107 |
| 重写版Tulu 3（Tulu 3 Rewritten，[论文](https://arxiv.org/abs/2411.15124)） | 7,109 |
| 多学科RLVR（Multi-Subject RLVR，[论文](https://arxiv.org/abs/2503.23829v1)） | 7,106 |
| AceReason-Math数据集（AceReason-Math，[论文](https://arxiv.org/abs/2505.16400)） | 6,598 |
| 英文WildChat（WildChat English，[论文](https://arxiv.org/abs/2405.01470)） | 6,421 |
| KlearReasoner代码数据集（KlearReasoner Code） | 6,272 |
| SYNTHETIC-2 / PrimeIntellect（[博客](https://www.primeintellect.ai/blog/synthetic-2)） | 3,000 |
| MathSub-30K（KlearReasoner数学数据集，[论文](https://arxiv.org/abs/2508.07629)） | 2,999 |
| ORZ数学数据集（ORZ Math，[论文](https://arxiv.org/abs/2503.24290)） | 2,999 |
| DAPO-Math数据集（DAPO-Math，[论文](https://arxiv.org/abs/2503.14476)） | 2,584 |
| Llama-Nemotron后训练数据集（Llama-Nemotron Post-Training Dataset，[论文](https://arxiv.org/abs/2505.00949)） | 2,006 |

### 分组混合数据集来源统计
| 混合数据集类型 | 样本数 |
|------|-------|
| 数学RLVR混合集（Math RLVR Mixture） | 30,180 |
| 指令遵循RLVR混合集（IF RLVR Mixture） | 29,813 |
| 代码RLVR混合集（Code RLVR Mixture） | 21,385 |
| 通用RLVR混合集（General RLVR Mixture） | 20,636 |

---

## 数据源与说明

### 指令遵循任务
- 最多包含5项约束条件
- 衍生自IFBench-Train与IFEval风格任务
- 针对清晰度与无毒性进行了筛选

### 数学推理任务
涵盖以下数据源：
- **OMEGA**
- **AceReason-Math**
- **ORZ Math**
- **DAPO-Math**
- **MathSub-30K**
覆盖几何、代数、组合数学、定理证明等广泛领域。

### 代码推理任务
包含四大类数据源：
- **AceCoder**
- **KlearReasoner-Code**
- **SYNTHETIC-2 / PrimeIntellect**
- **Llama-Nemotron后训练数据集**
所有数据源均通过测试用例执行验证进行筛选。

### 通用长文本推理任务
- 多学科RLVR数据集
- 重写版Tulu 3（通过F1分数进行筛选）
- 英文WildChat（针对推理适配性进行筛选）

---

## 处理与筛选流程

- 基于执行的代码筛选（通过测试用例验证）
- 面向安全性与质量的主题筛选
- 基于F1分数的重写结果筛选（针对Tulu 3）
- 按难度分层的Nemotron子集
- 严格去重处理
- 约束条件标准化

---

## 许可证
本数据集采用ODC-BY许可证发布，旨在遵循[AllenAI负责任使用指南](https://allenai.org/responsible-use)，供研究与教育场景使用。

## 引用格式

@misc{olmo2025olmo3,
title={Olmo 3},
author={Team Olmo and Allyson Ettinger and Amanda Bertsch and Bailey Kuehl and David Graham and David Heineman and Dirk Groeneveld and Faeze Brahman and Finbarr Timbers and Hamish Ivison and Jacob Morrison and Jake Poznanski and Kyle Lo and Luca Soldaini and Matt Jordan and Mayee Chen and Michael Noukhovitch and Nathan Lambert and Pete Walsh and Pradeep Dasigi and Robert Berry and Saumya Malik and Saurabh Shah and Scott Geng and Shane Arora and Shashank Gupta and Taira Anderson and Teng Xiao and Tyler Murray and Tyler Romero and Victoria Graf and Akari Asai and Akshita Bhagia and Alexander Wettig and Alisa Liu and Aman Rangapur and Chloe Anastasiades and Costa Huang and Dustin Schwenk and Harsh Trivedi and Ian Magnusson and Jaron Lochner and Jiacheng Liu and Lester James V. Miranda and Maarten Sap and Malia Morgan and Michael Schmitz and Michal Guerquin and Michael Wilson and Regan Huff and Ronan Le Bras and Rui Xin and Rulin Shao and Sam Skjonsberg and Shannon Zejiang Shen and Shuyue Stella Li and Tucker Wilde and Valentina Pyatkin and Will Merrill and Yapei Chang and Yuling Gu and Zhiyuan Zeng and Ashish Sabharwal and Luke Zettlemoyer and Pang Wei Koh and Ali Farhadi and Noah A. Smith and Hannaneh Hajishirzi},
year={2025},
eprint={2512.13961},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2512.13961},
}