ELIA-Qwen3-14B-LoRA

This repository contains the LoRA adapter weights for ELIA (Epilepsy Literature Intelligent Assistant), an epilepsy-specific literature question-answering assistant developed for evidence-grounded retrieval-augmented generation (RAG) workflows.

This repository is intended to store the adapter only. The Qwen3-14B base model is not included and should be obtained separately from its original repository.

Base model

• Base model: Qwen/Qwen3-14B
• Adapter type: LoRA
• Training framework: LLaMA-Factory / PEFT
• Task type: causal language modeling

If a different public base-model repository was used in your local training environment, update both this model card and adapter_config.json before release.

Intended use

ELIA is intended for research on epilepsy literature question answering and evidence-grounded clinical knowledge support. It is designed to be used together with an external retrieval module that supplies source evidence and citation metadata.

The model is not intended to provide independent medical diagnosis, treatment recommendations, or direct clinical decision-making. Outputs should be reviewed by qualified clinicians or domain experts, especially for high-risk clinical questions.

Limitations

• The adapter does not contain the retrieval index, full-text corpus, or PubMed/PDF archive.
• Performance depends on the retrieval system, evidence corpus, prompting strategy, and decoding configuration.
• The model may still generate incomplete, unsupported, or incorrect answers if retrieval fails or the question is outside the evaluated scope.
• Preprint-derived evidence should be clearly labeled and interpreted cautiously.

Training and evaluation data

The adapter was fine-tuned on epilepsy literature question-answering examples generated from the ELIA training pipeline. The full-text corpus and verbatim text chunks are not redistributed in this model repository because they may be subject to copyright and licensing restrictions.

During training, the following validation losses were recorded:

Epoch	Step	Validation loss
0.7407	100	0.3266
1.4815	200	0.3157
2.2222	300	0.3334
2.9630	400	0.3345

The lowest recorded validation loss was 0.3157 at step 200. Before final public release, ensure that adapter_model.safetensors corresponds to the selected final or best checkpoint.

Files in this repository

Expected release files include:

• adapter_model.safetensors
• adapter_config.json
• tokenizer.json
• tokenizer_config.json
• special_tokens_map.json
• added_tokens.json
• chat_template.jinja
• merges.txt
• vocab.json
• README.md

Intermediate checkpoints, local training logs, local paths, API tokens, raw full-text files, and retrieval indices should not be uploaded to this repository.

Example loading code

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen3-14B"
adapter_id = "YOUR_HF_USERNAME/ELIA-Qwen3-14B-LoRA"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, adapter_id)

Replace YOUR_HF_USERNAME/ELIA-Qwen3-14B-LoRA with the actual Hugging Face repository ID.

Citation

Please cite the associated ELIA manuscript when using this adapter. A formal citation will be added after publication.

License

This adapter is released under the Apache 2.0 license, subject to compatibility with the base model license and the terms of any underlying data sources used during development.