Yale-BIDS-Chen/LEME-Training-Data

LEME Training Data

Instruction and preference datasets for the LEME ophthalmology language models.

Paper · Code · Corpus

Motivation and data construction

The datasets were designed to adapt a general-purpose language model to ophthalmology: SFT develops domain-specific instruction following, while DPO explicitly targets truthful and informative responses.

For SFT, GPT-4o generated instruction-response examples from selected ophthalmology guidelines, StatPearls text, and published case reports, guided by clinician-designed prompts. The examples cover question answering, clinical information extraction, and patient-clinician dialogue, with response refinement and consistency checks to improve quality.

For DPO, prompts were sampled from the LEME instruction data and BioInstruct to cover ophthalmology and broader biomedical tasks. LEME-SFT responses were evaluated by GPT-4o for truthfulness and informativeness, then paired with revised responses to form preferred and dispreferred alternatives.

Configuration	Examples	Purpose	Model
sft	211,149	Supervised fine-tuning	LEME-SFT-70B
dpo	29,747	Direct preference optimization	LEME-DPO-70B

Both configurations use the train split.

Loading

from datasets import load_dataset

sft = load_dataset("Yale-BIDS-Chen/LEME-Training-Data", "sft", split="train")
dpo = load_dataset("Yale-BIDS-Chen/LEME-Training-Data", "dpo", split="train")

Data format

SFT: id identifies the example; conversations contains the user instruction and assistant response. Each message has from (human or gpt) and value (text) fields.

SFT task	Examples
Open-ended question answering	69,527
Multiple-choice question answering	84,248
USMLE-style question answering	10,842
Clinical information extraction	41,012
Patient–clinician dialogue	5,520

DPO: conversations contains the user prompt; chosen and rejected contain the preferred and alternative responses. Both responses use the same from and value message fields.

Contact

For general inquiries, please contact Hyunjae Kim at hyunjae.kim@yale.edu.

License

LEME code, models, data, and other related materials may be used for non-commercial research purposes. Any derivative works must be made publicly available through a public platform, such as GitHub or Hugging Face. For commercial-use inquiries, please contact us (agreements@yale.edu or qingyu.chen@yale.edu). Please see the LICENSE file for the full terms.

Source attribution and applicable third-party terms are listed in THIRD_PARTY_NOTICES.md.

Citation

@article{kim2026leme,
  title={LEME: open large language models for ophthalmology with advanced reasoning and clinical validation},
  author={Kim, Hyunjae and Ai, Xuguang and Srinivasan, Sahana and Gilson, Aidan and Singer, Maxwell B and Pushpanathan, Krithi and Xie, Qianqian and Park, Jungwoo and Applebaum, Serina and Yang, Gabriel Dawei and others},
  journal={npj Digital Medicine},
  year={2026},
  publisher={Nature Publishing Group UK, London}
}