kadirify/ medlocal-lab-explainer:latest

9 Downloads Updated 1 week ago

ollama run kadirify/medlocal-lab-explainer

curl http://localhost:11434/api/chat \
  -d '{
    "model": "kadirify/medlocal-lab-explainer",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='kadirify/medlocal-lab-explainer',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'kadirify/medlocal-lab-explainer',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 week ago

1 week ago

b4b27d3ffc20 · 4.3GB

model

archqwen3

·

parameters4.02B

·

quantizationQ8_0

4.3GB

params

{ "num_ctx": 4096, "stop": [ "<|im_end|>" ], "temperature": 0 }

65B

template

{{- range .Messages }}<|im_start|>{{ .Role }} {{ .Content }}<|im_end|> {{ end }}<|im_start|>assistan

102B

Readme

MedLocal Lab Explainer

A fine-tuned Qwen3-4B model that summarises laboratory report results in plain language for patient communication. Part of MedLocal, a clinical support tool for physicians.

This model is a research artifact. It is not a medical device. The clinical decision is always the physician’s.

What it does

Given a list of lab results with out-of-range flags and vetted background notes, the model writes a plain-language summary suitable for patient communication. It supports both English and Turkish.

The model does not decide whether a value is abnormal — that is done by deterministic code. It only restates vetted catalog notes in plain language.

Quick start

ollama pull kadirify/medlocal-lab-explainer

For the full application with PDF parsing, reference-range flagging, and a web interface, see the MedLocal repository.

Training details

Base: Qwen3-4B-Instruct-2507 (Apache 2.0)
Method: QLoRA (rank 16, alpha 16)
Data: 2,113 synthetic lab reports with teacher-generated summaries (Qwen3-14B), half Turkish
Hardware: Single RTX 5070 (12 GB), ~80 minutes
Quantisation: Q8_0 GGUF

Every training example passes eleven rule-based safety checks. The model is grounded: it receives the catalog’s vetted background for every flagged result and restates it, rather than recalling medicine from memory.

Evaluation (262 held-out synthetic reports)

	English	Turkish
Qwen3-4B, untuned	63.9%	20.0%
Fine-tuned	89.3%	92.1%

Rule compliance on synthetic data, not medical accuracy.

Limitations

Not clinically validated
Turkish phrasing can be stiff
No coverage for result combinations, only individual analytes
Reference ranges are general adult values

License

Apache 2.0

Links

# MedLocal Lab Explainer

A fine-tuned Qwen3-4B model that summarises laboratory report results in plain language for patient communication. Part of [MedLocal](https://github.com/kadirify/MedLocal), a clinical support tool for physicians.

> This model is a research artifact. It is not a medical device. The clinical decision is always the physician's.

## What it does

Given a list of lab results with out-of-range flags and vetted background notes, the model writes a plain-language summary suitable for patient communication. It supports both English and Turkish.

The model **does not decide** whether a value is abnormal — that is done by deterministic code. It only restates vetted catalog notes in plain language.

## Quick start

```bash
ollama pull kadirify/medlocal-lab-explainer
```

For the full application with PDF parsing, reference-range flagging, and a web interface, see the [MedLocal repository](https://github.com/kadirify/MedLocal).

## Training details

- **Base:** Qwen3-4B-Instruct-2507 (Apache 2.0)
- **Method:** QLoRA (rank 16, alpha 16)
- **Data:** 2,113 synthetic lab reports with teacher-generated summaries (Qwen3-14B), half Turkish
- **Hardware:** Single RTX 5070 (12 GB), ~80 minutes
- **Quantisation:** Q8_0 GGUF

Every training example passes eleven rule-based safety checks. The model is grounded: it receives the catalog's vetted background for every flagged result and restates it, rather than recalling medicine from memory.

## Evaluation (262 held-out synthetic reports)

| | English | Turkish |
|---|---|---|
| Qwen3-4B, untuned | 63.9% | 20.0% |
| **Fine-tuned** | **89.3%** | **92.1%** |

Rule compliance on synthetic data, not medical accuracy.

## Limitations

- Not clinically validated
- Turkish phrasing can be stiff
- No coverage for result combinations, only individual analytes
- Reference ranges are general adult values

## License

Apache 2.0

## Links

- [GitHub](https://github.com/kadirify/MedLocal)
- [Hugging Face (GGUF)](https://huggingface.co/kadirify/medlocal-lab-4b-gguf)

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)