---
language:
- en
license: apache-2.0
base_model: LiquidAI/LFM2.5-230M
tags:
- medical
- qat
- lora
- torchao
- trl
- sft
model_name: vilyalabs-med
---

# vilyalabs-med

This model is a fine-tuned version of [LiquidAI/LFM2.5-230M](https://huggingface.co/LiquidAI/LFM2.5-230M) optimized for medical chat and consultation. 

It was trained using **Quantization-Aware Training (QAT)** for INT4 weight-only quantization via `torchao`, combined with **LoRA** adapters to maintain high reasoning quality while being extremely lightweight.

## Model Details
- **Base Model:** LiquidAI/LFM2.5-230M
- **Fine-tuning Method:** LoRA + INT4 QAT (Weight-only)
- **LoRA Config:** rank=16, alpha=32, target_modules=all-linear
- **Quantization:** INT4 Weight-Only (Group size 128)
- **Training Precision:** FP16

## Training Data
The model was trained on a curated mixture of:
1. `ruslanmv/ai-medical-chatbot`
2. `lavita/ChatDoctor-HealthCareMagic-100k` 

## Intended Use
- Medical information retrieval
- Patient-Doctor dialogue simulation
- Health-related assistant tasks

## How to use
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("micymike/vilyalabs-med", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("micymike/vilyalabs-med")
```