Magicoder-OSS-Instruct-75K
Uses real open-source code snippets as seeds to generate diverse coding problems and solutions with GPT-4. The OSS-Instruct method produces more realistic, grounded coding tasks than purely synthetic approaches. Magicoder-S-DS-6.7B trained on this data surpassed GPT-3.5-Turbo on HumanEval.
from datasets import load_dataset
ds = load_dataset("ise-uiuc/Magicoder-OSS-Instruct-75K")Preview a sample row
{
"problem": "Given a list of integers, write a function to find all pairs that sum to a target value. Return pairs as tuples sorted by first element.",
"solution": "def find_pairs(nums: list[int], target: int) -> list[tuple[int, int]]:\n seen = set()\n pairs = []\n for n in nums:\n complement = target - n\n if complement in seen:\n pairs.append((min(n, complement), max(n, complement)))\n seen.add(n)\n return sorted(set(pairs))"
}Fine-tune with this dataset
Estimated VRAM to fine-tune with QLoRA (4-bit base model + LoRA adapters), using conservative defaults:
Frequently asked questions
Can I use Magicoder-OSS-Instruct-75K commercially?
Yes — Magicoder-OSS-Instruct-75K is released under MIT, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does Magicoder-OSS-Instruct-75K contain, and do I need all of it?
Magicoder-OSS-Instruct-75K contains 75k Rows. You rarely need all of it: for style and format fine-tuning, a few hundred to a few thousand examples are enough — load a slice (e.g. split="train[:1000]") and scale up only if quality plateaus.
What is Magicoder-OSS-Instruct-75K best used for?
Code instruction tuning grounded in real open-source snippets. It belongs to the Code section of our dataset hub, where you'll find alternatives and complementary sets.