README.md

---
license: apache-2.0
language:
- en
tags:
- reasoning
- instruction-tuning
- chatml
- math
- coding
- general
- orbit
pretty_name: Orbit-200K
size_categories:
- 100K<n<1M
task_categories:
- text-generation
configs:
- config_name: default
  data_files:
  - split: train
    path: "data/train-00000-of-00001.parquet"
dataset_info:
  features:
  - name: text
    dtype: string
  splits:
  - name: train
    num_examples: 200000
  download_size: 168000000
  dataset_size: 200000000
---

# 🪐 Orbit-200K

<p align="center">
  <img src="./banner.png" width="100%">
</p>

> **A high-signal instruction dataset designed for training universal language models with dense reasoning, coding, mathematics, and general intelligence—without conversational bloat.**

Orbit-200K is a carefully curated dataset of **200,000 instruction-response pairs** optimized for training modern language models ranging from **0.5B to 7B parameters**.

Unlike many public instruction datasets, Orbit-200K removes unnecessary conversational filler and focuses on maximizing useful learning signal per token, making it especially suitable for efficient fine-tuning of compact models.

---

# Overview

| Property | Value |
|----------|-------|
| Examples | 200,000 |
| Language | English |
| Format | ChatML |
| License | Apache-2.0 |
| File Format | Parquet |
| Column | `text` |
| Training Target | Instruction Fine-tuning |
| Recommended Models | Qwen, Llama 3, Gemma, Mistral, Phi |

---

# Why Orbit-200K?

Most instruction datasets contain significant amounts of unnecessary dialogue:

- "Certainly!"
- "Let's break this down."
- "I'd be happy to help."
- "I hope this helps!"

While these phrases improve user experience, they contribute very little to model capability and consume valuable context length.

Orbit-200K removes this conversational overhead so the model learns to produce:

- direct answers
- concise reasoning
- efficient coding solutions
- dense mathematical explanations

This results in higher information density during training.

---

# Key Features

## 🚀 Zero Conversational Bloat

Responses begin immediately with useful content rather than assistant preambles.

Instead of:

> Certainly! Let's solve this step by step...

the model learns outputs like:

> Compute the derivative using the product rule...

---

## 🧠 Balanced Instruction Mixture

Orbit-200K combines three complementary domains:

| Dataset Source | Percentage | Purpose |
|----------------|-----------:|---------|
| MetaMathQA | 37.5% | Mathematical reasoning and logical thinking |
| OSS-Instruct (Exec Verified) | 25% | Coding, algorithms, software engineering |
| OpenHermes-2.5 (Filtered) | 37.5% | General knowledge, instruction following, QA |

This mixture provides strong general-purpose instruction tuning while maintaining excellent reasoning capability.

---

## 💬 Native ChatML Formatting

Every sample is already formatted using ChatML.

Compatible with:

- Qwen
- Llama 3
- Gemma
- Mistral
- most ChatML-based instruction models

No additional preprocessing is required.

---

# Dataset Structure

The dataset contains a single column.

| Column | Type | Description |
|---------|------|-------------|
| text | string | Fully formatted ChatML conversation |

Example:

```text
<|im_start|>system
You are a helpful, logical, and precise AI assistant.
<|im_end|>

<|im_start|>user
Write a Python function to check if a string is a palindrome.
<|im_end|>

<|im_start|>assistant
def is_palindrome(s):
    return s == s[::-1]
<|im_end|>
```

---

# Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset(
    "Siddh07ETH/Orbit-200K",
    split="train"
)
```

Example:

```python
print(dataset[0]["text"])
```

---

# Data Processing Pipeline

Orbit-200K underwent multiple quality-control stages.

## 1. Conversational De-bloating

Removed unnecessary assistant phrases such as:

- Certainly!
- Of course!
- I'd be happy to help!
- Let's solve this together.
- I hope this helps!

---

## 2. Length Filtering

Very short assistant responses were removed.

Minimum assistant response length:

> **50 characters**

This encourages denser learning signals.

---

## 3. Exact Duplicate Removal

Duplicate instruction-response pairs were removed to maximize unique information.

---

## 4. Formatting

All samples were converted into standardized ChatML format.

---

# Recommended Use Cases

Orbit-200K is suitable for:

- Instruction tuning
- Supervised Fine-Tuning (SFT)
- Small language models (0.5B–7B)
- Reasoning models
- Coding assistants
- Mathematical reasoning
- General-purpose chat models

---

# Recommended Training Models

Orbit-200K works particularly well with:

- Qwen 2.5
- Qwen 3
- Llama 3
- Gemma 2
- Mistral
- Phi

---

# Training Objective

The dataset is designed to improve:

- Instruction following
- Logical reasoning
- Mathematical problem solving
- Programming capability
- Response density
- Reduced verbosity

---

# Statistics

| Metric | Value |
|---------|-------|
| Total Examples | 200,000 |
| Number of Columns | 1 |
| Primary Column | `text` |
| Format | ChatML |
| File Format | Parquet |
| License | Apache-2.0 |

---

# Citation

If you use Orbit-200K in your work, please cite:

```bibtex
@dataset{orbit200k2026,
  title={Orbit-200K: A High-Signal Instruction Dataset for Universal LLM Reasoning},
  author={Siddharth},
  year={2026},
  publisher={Hugging Face}
}
```

---

# License

This dataset is released under the **Apache License 2.0**.

See the LICENSE file for details.

---

# Acknowledgements

Orbit-200K builds upon publicly available datasets, including:

- MetaMathQA
- OSS-Instruct (Exec Verified)
- OpenHermes-2.5

Credit belongs to the original dataset creators whose work made this curated mixture possible.

---

# Author

**Siddharth**

Hugging Face: **https://huggingface.co/Siddh07ETH**

---

### ⭐ If Orbit-200K helps your research or model training, consider giving the dataset a Like on Hugging Face.