README.md

---
license: apache-2.0
language:
- en
tags:
- distillation
- coding
- agentic
- kimi-k3
- gpt-5.6
- fable-5
- frontier-models
- chatml
- instruction-tuning
- code-generation
pretty_name: Atlas-Frontier-Model-Traces
size_categories:
- 10K<n<100K
task_categories:
- text-generation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: text
    dtype: string
  - name: system_prompt
    dtype: string
  - name: source_model
    dtype: string
  splits:
  - name: train
    num_bytes: 66731283
    num_examples: 15746
  download_size: 36083093
  dataset_size: 138631283
---

# 🧬 Atlas-Frontier-Model-Traces

A universal ChatML dataset distilling the agentic coding capabilities of frontier models (Kimi-K3, GPT-5.6-Sol, and Fable-5).

---

# Dataset Description

**Atlas-Frontier-Model-Traces** is a meticulously curated dataset containing **15,746 coding and debugging traces** generated by three of the most advanced frontier AI models.

This dataset is designed for **Knowledge Distillation**. By training smaller open-source language models (such as **Qwen, Llama, Mistral, and Gemma**) on the output traces of frontier-scale models, developers can transfer sophisticated coding behaviors into efficient edge-deployable models.

---

# Data Sources

This dataset is created by combining and rigorously cleaning the following public datasets:

- `greghavens/kimi-k3-coding-and-debugging-traces`
- `greghavens/gpt-5.6-sol-coding-and-debugging-traces`
- `greghavens/fable-5-coding-and-debugging-traces`

---

# Dataset Structure

The dataset has been converted into a universal ChatML format compatible with virtually every modern LLM architecture.

| Column | Type | Description |
|---------|------|-------------|
| `text` | string | ChatML conversation containing only user and assistant turns |
| `system_prompt` | string | Suggested system prompt stored separately for easier customization |
| `source_model` | string | Metadata indicating the originating frontier model |

---

# Example Entry

```text
<|im_start|>user
Write a Python function to connect to a PostgreSQL database.
<|im_end|>
<|im_start|>assistant
import psycopg2

def connect_to_db(dbname, user, password, host, port):
    # Function implementation here...
<|im_end|>
```

---

# Data Processing & Cleaning Pipeline

To maximize quality and eliminate schema inconsistencies, the dataset was processed using a robust MLOps pipeline.

- **Pandas Bypass Loading**
  - Loaded raw datasets with Pandas to avoid Hugging Face schema conflicts across repositories.

- **Aggressive Schema Unification**
  - Converted ShareGPT, OpenAI Messages, Prompt/Completion, and other formats into one standardized schema.

- **Quality Filtering**
  - Removed broken samples and assistant responses shorter than 15 characters.

- **Universal ChatML Formatting**
  - Extracted hardcoded system prompts from conversations and stored them in a separate column.

- **Parquet Compression**
  - Reduced over **1 GB** of raw data into a compact **36.1 MB** high-signal Parquet dataset.

---

# How to Use

```python
from datasets import load_dataset

# Load dataset
ds = load_dataset(
    "Siddh07ETH/Atlas-Frontier-Model-Traces",
    split="train"
)

custom_sys_prompt = "You are a helpful coding assistant."

def format_for_training(example):
    return {
        "final_text":
        f"<|im_start|>system\n"
        f"{custom_sys_prompt}"
        f"<|im_end|>\n"
        f"{example['text']}"
    }

ds = ds.map(format_for_training)

print(ds[0]["final_text"])
```

---

# Intended Uses

This dataset is intended for:

- Knowledge Distillation
- Instruction Tuning
- Supervised Fine-Tuning (SFT)
- Coding Assistants
- Agentic Tool Use
- Debugging Models
- Fine-tuning 0.5B–7B parameter language models

Supported architectures include:

- Qwen
- Llama
- Mistral
- Gemma
- Other ChatML-compatible models

---

# Limitations

- Focused exclusively on coding and debugging tasks.
- Does not contain general conversational or multilingual data.
- Optimized for functional code generation and debugging rather than long-form software architecture discussions.

---

# Citation

If you use this dataset in your research or projects, please cite both this dataset and the original source datasets where appropriate.

---

# License

This dataset is distributed under the **Apache 2.0 License**, inheriting the licensing terms of the original source datasets.