📚 Atlas-Coder-50K-ChatML

A 50,000-sample execution-verified Python instruction dataset, pre-formatted in ChatML for immediate supervised fine-tuning of coding language models.

This dataset accompanies Atlas-Coder-2-0.5B and represents the exact high-quality training subset used during instruction tuning.

---

✨ Highlights

• ✅ 50,000 curated Python instruction-response pairs
• ✅ Execution-verified (compiled and tested before inclusion)
• ✅ Pre-formatted in ChatML
• ✅ Ready for Hugging Face Trainer / SFTTrainer
• ✅ Decontaminated against HumanEval & MBPP
• ✅ Optimized for 0.5B–3B parameter coding models

---

📖 Dataset Description

Atlas-Coder-50K-ChatML is a carefully processed subset of the bigcode/self-oss-instruct-sc2-exec-filter-50k dataset.

Rather than increasing dataset size, the focus is on maximizing signal quality, enabling smaller language models to efficiently learn Python programming, algorithmic reasoning, and instruction following.

Each sample is already formatted in standard ChatML, eliminating the need for additional preprocessing before fine-tuning.

---

📂 Dataset Format

The dataset contains a single column:

Column	Type	Description
text	string	Complete ChatML conversation

Example:

<|im_start|>system
You are a helpful coding assistant.
<|im_end|>
<|im_start|>user
Write a function that returns the sum of all even numbers.
<|im_end|>
<|im_start|>assistant
def sum_even_numbers(lst):
    return sum(x for x in lst if x % 2 == 0)
<|im_end|>

---

🚀 Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "Siddh07ETH/Atlas-Coder-50K-ChatML",
    split="train"
)

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

---

🛡️ Benchmark Integrity

Before training Atlas-Coder-2, the dataset was decontaminated against:

• HumanEval
• MBPP

using an 8-gram Jaccard similarity filter with a 0.30 overlap threshold.

Samples with excessive similarity to benchmark prompts were removed to preserve fair evaluation and prevent benchmark leakage.

---

🎯 Intended Use

This dataset is designed for:

• Instruction tuning
• Supervised fine-tuning (SFT)
• QLoRA training
• Small Language Models (SLMs)
• Python code generation
• Code completion
• Algorithmic reasoning

Recommended model sizes:

• 0.5B
• 1B
• 1.5B
• 3B

---

⚠️ Limitations

• Python-only dataset
• Does not include multilingual programming languages
• Optimized for functional code generation rather than long-form software engineering discussions

---

📄 License

This dataset is released under the Apache-2.0 license and follows the licensing terms of the original bigcode/self-oss-instruct-sc2-exec-filter-50k dataset.