TNSA/OpenReasoner-V1

---
license: mit
language:
- en
metrics:
- accuracy
tags:
- reasoning
- math
- instruction-tuning
- ngen
---


# OpenReasoner-V1

OpenReasoner-V1 is a high-fidelity, unified dataset designed for fine-tuning advanced reasoning models. It combines complex mathematical problem-solving with high-quality general instruction data, specifically optimized for state-of-the-art small language models like **NGen-4 Lite**.

## 🚀 Key Features

- **Integrated Math Reasoning**: Includes a curated selection of 220k+ mathematical problems with detailed, step-by-step solutions to improve logical deduction capabilities.
- **Hermes-Core Alignment**: Infused with the OpenHermes-2.5 instruction set to ensure the model maintains excellent general-purpose conversational abilities and broad knowledge.
- **Deep-Thinking Optimization**: Features distilled reasoning trajectories formatted with `<think>` tags, designed to teach models how to "reason-before-acting" in complex scenarios.
- **Unified Format**: All data is provided in a standard multi-turn conversation format, ready for immediate use in SFT (Supervised Fine-Tuning) pipelines.

## 📊 Data Composition

| Component | Source Type | Focus |
| :--- | :--- | :--- |
| **Math-Logic** | OpenR1-Math | Theorem proving, calculus, and multi-step logic. |
| **General Instruction** | OpenHermes-2.5 | Creativity, coding, and general knowledge. |
| **Distilled Reasoning** | High-Fidelity SOTA Distillation | Advanced "Chain of Thought" (CoT) and strategy. |

## 🛠 Usage

This dataset is ideal for training models that require a balance between **high-IQ reasoning** and **low-latency instruction following**. It is the primary training source for the **NGen-4.1** series.

### Formatting
Reasoning samples follow the established `<think>` pattern: