Cyleux/

gemma3-reasoning-dropin

Drop-in replacement dataset for Gemma3-style notebooks expecting task, input, expected_output fields.

from datasets import load_dataset

dataset = load_dataset("Cyleux/gemma3-reasoning-dropin", split="train[:10000]")

Then your existing mapping code can stay the same:

task -> system input -> user expected_output -> assistant

Reasoning is preserved in expected_output using ... blocks.

{ "input":… See the full description on the dataset page: https://huggingface.co/datasets/Cyleux/gemma3-reasoning-dropin.