---
license: cc-by-4.0
language:
- en
library_name: peft
pipeline_tag: text-generation
base_model: Qwen/Qwen3-1.7B-Base
base_model_relation: finetune
datasets:
- luizaaca/symptoms-to-diseases-with-reasoning
tags:
- medical
- clinical-screening
- symptom-analysis
- structured-output
- json
- lora
- qlora
- gguf
- llama.cpp
- ollama
- unsloth
- peft
- transformers
---

# Qwen3-1.7B Clinical Screening

This repository packages the artifacts generated by the training notebook
[`screening_robot_qwen3_1_7b_json.ipynb`](https://github.com/luizaaca/screening_robot/blob/main/screening_robot_qwen3_1_7b_json.ipynb) for the `symptom_analysis`
specialist step of the Screening Robot workflow.

## Artifact layout

- `lora/`: PEFT LoRA adapter weights and tokenizer/chat-template files.
- `gguf/qwen3-1.7b-clinical-screening.Q4_K_M.gguf`: merged and quantized GGUF export (`Q4_K_M`) for
  llama.cpp and Ollama-style runtimes.
- `gguf/config.json`, `gguf/tokenizer.json`, `gguf/tokenizer_config.json`,
  `gguf/chat_template.jinja`: support files exported alongside the GGUF build.
- `Modelfile`: Ollama-oriented template aligned with the same JSON-focused
  system prompt used in the training notebook.

## Model details

- **Repository**: `https://huggingface.co/luizaaca/qwen3-1.7b-clinical-screening`
- **Base model**: `Qwen/Qwen3-1.7B-Base`
- **Training runtime base**: `unsloth/qwen3-1.7b-base-unsloth-bnb-4bit`
- **Training recipe**: QLoRA via Unsloth on a 4-bit loaded base model
- **LoRA hyperparameters**: rank 16, alpha 32, dropout 0.0
- **Target modules**: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`,
  `up_proj`, `down_proj`
- **Max sequence length used during training**: 1024
- **Training runtime**: Google Colab + Tesla T4 (16 GB VRAM)
- **Kaggle Dataset**: [`luizaaca/symptoms-to-diseases-with-reasoning`](https://www.kaggle.com/datasets/luizaaca/symptoms-to-diseases-with-reasoning)

## Intended behavior

The model was fine-tuned to emit a compact JSON object with exactly three keys:

- `support_status`
- `candidate_diseases`
- `recommended_exams_tests`

The fine-tuning notebook explicitly scoped the model to the
dataset-aligned specialist contract used by the `symptom_analysis` node.
It does **not** represent the full downstream production payload used
elsewhere in the application.

## Prompt contract

Training and validation were built around the following behavior:

- a clinical screening system prompt focused on symptom analysis;
- a user message formatted as `Clinical request` plus `Active patient context`; and
- an assistant answer restricted to JSON only.

The default `Modelfile` included in this repository mirrors that setup.

## Validation summary

The notebook includes:

- a held-out evaluation pass comparing the fine-tuned model against the base model;
- an assertion that fine-tuned label accuracy matches or exceeds the base model on the evaluation split; and
- GGUF smoke tests that require the exported model to emit valid schema-compliant JSON.

No claim of clinical validation, diagnostic safety, or regulatory readiness is made.

## Notes about the GGUF filename

The local GGUF export produced by the toolchain may inherit the original base
model filename. In this repository the uploaded GGUF weight is renamed to
`qwen3-1.7b-clinical-screening.Q4_K_M.gguf` to make it explicit that the file contains the
fine-tuned clinical screening variant.

## Intended use

This repository is suitable for:

- research experiments on structured clinical-screening assistants;
- integration prototypes for symptom-intake workflows; and
- local inference with PEFT or GGUF-compatible runtimes.

## Out-of-scope use

This repository is **not** intended for:

- autonomous diagnosis or treatment decisions;
- emergency or high-acuity triage without clinician oversight;
- prescribing or medication guidance; or
- use as a substitute for professional medical judgment.

## Safety notice

This is a research artifact for screening-assistance workflows only. Always
keep a qualified human clinician in the loop.