ealharbi/crystallography-cryoem-mistral-large-123b-lora

Mistral Large 123B — crystallography LoRA adapter

LoRA adapter for mistralai/Mistral-Large-Instruct-2407, fine-tuned on confirmed question-and-answer pairs from the CCP4BB, CCP-EM, phenixbb and COOT mailing lists.

Released under the Mistral Research Licence: research and non-commercial use only, inherited from the base model.

Training

Method	LoRA, r=64, α=128, dropout=0.05
Learning rate	2.675e-05
Max length	2048 tokens
Epochs	2, early stopping
Seed	1

Evaluation

BERTScore F1 against held-out confirmed answers (433 questions, roberta-large, score(prediction, reference)): 0.8320 → 0.8570.

BERTScore measures overlap with the reference wording, not factual correctness; the paper reports three LLM judges alongside it.

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("mistralai/Mistral-Large-Instruct-2407")
model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-Large-Instruct-2407", device_map="auto", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "ealharbi/crystallography-cryoem-mistral-large-123b-lora")

The system prompt used in training was:

You are a helpful crystallography assistant. Be concise and precise.

The training pairs are single-turn, so ask one self-contained question at a time; multi-turn prompts fall outside the fine-tuning format.

Caveat

Answers are generated and may be wrong. Verify anything consequential against the program documentation and the primary literature.