netdur/gemma-3-4b-radiology

Uploaded finetuned model

This repository contains a Gemma 3 4B model fine-tuned to generate radiology reports from ultrasound images. It was trained using the Unsloth library and is provided here in GGUF format for use with llama.cpp.

This model card demonstrates a complete, end-to-end example of running inference using llama.cpp and the llama_cpp_dart package.

• Base Model: unsloth/gemma-3-4b-it-unsloth-bnb-4bit
• Fine-tuning Data: unsloth/Radiology_mini

Example Usage

This example uses the llama_cpp_dart package to run the model on a macOS machine with Metal GPU acceleration.

1. Input Image

The following ultrasound image (radiology.png) was used as input.

2. Dart Inference Code

The model was called using the code below. Note the use of ChatFormat.gemma, which correctly applies the <bos><start_of_turn>user... template required by the model.

import 'dart:io';
import 'package:llama_cpp_dart/llama_cpp_dart.dart';

Future<void> main() async {
  Llama.libraryPath = "bin/MAC_ARM64/libmtmd.dylib";

  final modelParams = ModelParams()..nGpuLayers = -1;

  final contextParams = ContextParams()
    ..nPredict = 512
    ..nCtx = 8192
    ..nBatch = 8192;

  final samplerParams = SamplerParams()
    ..temp = 0.0
    ..topK = 64
    ..topP = 0.95
    ..penaltyRepeat = 1.1
    ..addStopSequence("<end_of_turn>");

  final llama = Llama(
      "./model-radiology-Q4_K_M.gguf",
      modelParams,
      contextParams,
      samplerParams,
      false,
      "./mmproj-radiology.gguf");

  final image =
      LlamaImage.fromFile(File("./radiology.png"));

  final chat = ChatHistory();
  chat.addMessage(role: Role.user, content: """<image>
      You are an expert radiographer. Describe accurately what you see in this image.""");
  
  // Use the correct chat format that matches the fine-tuning process
  final prompt =
      chat.exportFormat(ChatFormat.gemma, leaveLastAssistantOpen: true);

  print("==== PROMPT SENT TO MODEL ====");
  print(prompt);
  print("==============================");

  final sw = Stopwatch()..start();
  try {
    final stream = llama.generateWithMeda(prompt, inputs: [image]);

    await for (final token in stream) {
      stdout.write(token);
    }
    await stdout.flush();
    stdout.writeln();
  } on LlamaException catch (e) {
    stderr.writeln("An error occurred: $e");
  } finally {
    sw.stop();
    stdout.writeln('⏱️  Inference time: ${sw.elapsed}');
    llama.dispose();
  }
}

Outputs

• Image Type: Transvaginal ultrasound
• Date: 10-05-2012
• Measurement: 40mm
• Findings:
  • A large, heterogeneous, cystic mass is seen in the right adnexa.
  • The mass appears to be connected to the ovary.
  • The mass is 40mm in diameter.
  • The right ovary is 62.7 mm in diameter.
  • The left ovary is 125/255 mm in diameter.
  • The uterus is 3D5-8EK.
  • The uterine body is 11:36:15.
  • The uterine tube is 0.2.