metadata

license: apache-2.0
datasets:
  - harishnair04/mtsamples
language:
  - en
base_model:
  - google/gemma-2-2b
tags:
  - trl
  - sft
  - quantization
  - 4bit
  - LoRA
library_name: transformers

Model Card for Medical Transcription Model (Gemma-MedTr)

This model is a fine-tuned variant of Gemma-2-2b, optimized for medical transcription tasks with efficient 4-bit quantization and Low-Rank Adaptation (LoRA). It handles transcription processing, keyword extraction, and medical specialty classification.

Model Details

Developed by: Harish Nair
Organization: University of Ottawa
License: Apache 2.0
Fine-tuned from: Gemma-2-2b
Model type: Transformer-based language model for medical transcription processing
Language(s): English

Training Details

Training Loss: Final training loss at step 10: 1.4791
Training Configuration:
- LoRA with r=8, targeting specific transformer modules for adaptation.
- 4-bit quantization using nf4 quantization type and bfloat16 compute precision.
Training Runtime: 20.85 seconds, with approximately 1.92 samples processed per second.

How to Use

To load and use this model, initialize it with the following configuration:

import pandas as pd
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, PeftModel

model_id = "harishnair04/Gemma-medtr-2b-sft"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_id, token=access_token_read)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map='auto', token=access_token_read)