sft_data_mlx.py

# /// script
# requires-python = ">=3.11,<3.12"
# dependencies = [
#     "distilabel[mlx]",
# ]
# ///
from distilabel.models import MlxLLM
from distilabel.pipeline import InstructionResponsePipeline

llm = MlxLLM(
    path_or_hf_repo="mlx-community/Qwen2.5-32B-Instruct-4bit",
    use_magpie_template=True,
    magpie_pre_query_template="qwen2",
    generation_kwargs={"temp": 1, "max_tokens": 4000},
)

pipeline = InstructionResponsePipeline(llm=llm, batch_size=5, num_rows=10)

if __name__ == "__main__":
    dataset = pipeline.run()
    dataset.push_to_hub("davidberenstein1957/sft-dataset")

LinkedIn Post: https://www.linkedin.com/posts/davidberenstein_synthetic-data-for-fine-tuning-and-evaluating-activity-7287026349846781953-CI5G