Used mlx-community/Phi-3-mini-4k-instruct-4bit-no-q-embed to generate summaries from akemiH/MedQA-Reason:
import os
import datasets
from mlx_lm import load, generate
def _summarize(example):
prompt = f"<|user|>\n{example['input'].strip()}\n{example['output_reason']}\n\nSummarize the keypoint of the above question-answer pair into one sentence.<|end|>\n<|assistant|>"
example['summary'] = generate(model, tokenizer, prompt, max_tokens=500)
return example
model, tokenizer =… See the full description on the dataset page:
https://huggingface.co/datasets/JosefAlbers/akemiH_MedQA_Reason.
