This repository contains a LoRA adapter fine-tuned on the AIR-Bench Healthcare Spanish (ES) benchmark using Meta Llama 3.3 70B Instruct as the base model. The adapter is intended for research on healthcare-oriented instruction following, medical question answering, and clinical reasoning tasks in Spanish.

Developed by: Roshan G
Funded by: Independent research
Shared by: Roshan G
Model type: PEFT LoRA Adapter
Language(s): Spanish (es)
License: Subject to the license of the base model (Meta Llama 3.3)
Finetuned from model: meta-llama/Llama-3.3-70B-Instruct

Model Sources

Repository: https://huggingface.co/Roshang09112007/adaption-air-bench-healthcare-es
Base Model: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct

Uses

Direct Use

This model is intended for:

Medical question answering
Healthcare conversational systems
Clinical reasoning research
Instruction-following evaluation
Spanish medical NLP research

Downstream Use

Potential downstream applications include:

Medical chatbots
Clinical decision support research
Healthcare benchmark evaluation
Educational and academic projects
Domain adaptation studies

Out-of-Scope Use

This model is not intended for:

Real-world clinical diagnosis
Medical treatment recommendations
Autonomous healthcare decision making
Emergency medical advice
Any safety-critical healthcare application

Bias, Risks, and Limitations

This model inherits limitations from both:

The underlying Llama 3.3 70B Instruct model
The AIR-Bench Healthcare training data

Potential risks include:

Generation of incorrect medical information
Hallucinated medical facts
Biases present in healthcare datasets
Outdated clinical knowledge
Unsafe recommendations if used without expert oversight

Outputs should always be verified by qualified healthcare professionals.

Recommendations

Users should:

Treat outputs as research artifacts only.
Verify all medical information independently.
Avoid deploying this model in production healthcare systems.
Evaluate performance thoroughly before downstream usage.

How to Get Started with the Model

Install dependencies:

pip install transformers peft accelerate torch

Load the adapter:

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

BASE_MODEL = "meta-llama/Llama-3.3-70B-Instruct"
ADAPTER = "Roshang09112007/adaption-air-bench-healthcare-es"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER
)

prompt = "¿Cuáles son los síntomas de la diabetes tipo 2?"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=200
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

The model was fine-tuned using the AIR-Bench Healthcare Spanish (ES) dataset consisting of healthcare-related instruction-response pairs.

Training Procedure

The model was trained using Parameter-Efficient Fine-Tuning (PEFT) with the LoRA method.

Training Hyperparameters

Training regime: Supervised fine-tuning (SFT)
Fine-tuning method: LoRA
Framework: PEFT
Precision: Mixed precision
Base model: Llama 3.3 70B Instruct

Evaluation

Testing Data

Evaluation was performed using healthcare-related Spanish benchmark examples.

Factors

Evaluation considered:

Instruction following
Medical reasoning
Response relevance
Spanish language quality

Metrics

Potential evaluation metrics include:

Accuracy
Exact Match (EM)
BLEU
ROUGE
Human evaluation

Results

Evaluation results are currently under investigation and may vary depending on task and prompting strategy.

Summary

This repository provides a LoRA adapter specialized for healthcare-related Spanish instruction-following tasks using Meta Llama 3.3 70B Instruct as the backbone model.

Environmental Impact

Carbon emissions were not formally measured.

Hardware Type: GPU Accelerator
Hours Used: Not reported
Cloud Provider: Not reported
Compute Region: Not reported
Carbon Emitted: Unknown

Technical Specifications

Model Architecture and Objective

Architecture: Transformer Decoder
Base Model: Meta Llama 3.3 70B Instruct
Adaptation Method: LoRA (PEFT)
Objective: Instruction fine-tuning for healthcare tasks

Compute Infrastructure

Fine-tuning performed using GPU-based hardware acceleration.

Hardware

GPU: Not specified

Software

Python 3.10
Transformers
PEFT 0.15.1
PyTorch
Accelerate

Citation

BibTeX

bibtex

@misc{roshan2026airbenchhealthcare,
  title={AIR-Bench Healthcare ES LoRA Adapter},
  author={Roshan G},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Roshang09112007/adaption-air-bench-healthcare-es}}
}

APA

Roshan G. (2026). AIR-Bench Healthcare ES LoRA Adapter. Hugging Face.

More Information

This model is released for research and educational purposes only.

Model Card Authors

Roshan G

Model Card Contact

Hugging Face: https://huggingface.co/Roshang09112007

Framework Versions

PEFT: 0.15.1
Transformers: Compatible
PyTorch: Compatible