---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
library_name: peft
pipeline_tag: text-generation
tags:
- medical
- healthcare
- lora
- qlora
- peft
- calibration
- selective-prediction
---

# Qwen3-4B Medical Reasoning (QLoRA adapter)

**Title:** Qwen3-4B Medical Reasoning (QLoRA)
**Subtitle:** A QLoRA adapter fine-tuning Qwen3-4B-Instruct on medical chain-of-thought reasoning, trained on a single free Kaggle T4 - the SFT checkpoint from the "Confidence Tax" study.

> Reusable as both the Kaggle **model card** and the Medium/blog blurb.
> ⚠️ Research & education only - **NOT for clinical or patient-facing use.**

## Model Summary
A **LoRA adapter** that fine-tunes [`Qwen/Qwen3-4B-Instruct-2507`](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)
on **medical chain-of-thought reasoning**. It is the supervised-fine-tuned ("after") checkpoint from a
controlled **calibration study** of small medical LLMs. The headline finding: this fine-tune makes the
model's **internal** (token-probability) confidence markedly **better calibrated**, while its
**verbalized** (self-reported) confidence does not improve - the two diverge. Trained entirely on a
**single free Kaggle T4**.

## Model Characteristics
- **Base:** Qwen3-4B-Instruct-2507 (Apache-2.0), 4B params, text-only.
- **Method:** QLoRA - 4-bit NF4 (double-quant), LoRA **r=16, α=32, dropout 0.05** on
 `q/v/o/gate/up/down` projections. Trainable params well under 1% of the model.
- **Training:** 500 steps (~1.33 epochs), seq len 512, effective batch 16, lr 2e-4, fp16.
 Final/avg `train_loss = 1.477` (from ~1.7), grad-norm stable ~0.27, no divergence. ~3.5 h on one T4.
- **Format:** PEFT LoRA adapter (`adapter_config.json` + `adapter_model.safetensors`). Not a standalone model - 
 load on top of the base. (Base tokenizer is used; not bundled.)

## Data Overview
- [`FreedomIntelligence/medical-o1-reasoning-SFT`](https://huggingface.co/datasets/FreedomIntelligence/medical-o1-reasoning-SFT)
 (Apache-2.0): chain-of-thought medical Q&A, GPT-4o-generated reasoning validated by a medical verifier.
 **No PHI.** ~6,000 examples used.

## Evaluation Results
Before (base) vs. after (this adapter) on **MedQA + MedMCQA** (0-shot, n≈800), with **95% paired bootstrap
CIs** (2000 resamples). **Bold** = the change is statistically significant (CI excludes 0).

**Internal confidence** (max-softmax over option logits - the deployable signal):

| metric | base | +this adapter | Δ [95% CI] |
|---|---|---|---|
| Accuracy ↑ | 0.611 | 0.576 | **-0.036 [-0.063, -0.007]** |
| ECE ↓ (calibration error) | 0.351 | **0.085** | **-0.266 [-0.287, -0.220]** |
| AURC ↓ (selective prediction) | 0.223 | 0.236 | +0.013 [-0.013, +0.031] (n.s.) |
| Abstention rate | 0.125 | 0.086 | -0.039 [-0.060, -0.019] |

**Verbalized confidence** (the "Confidence: NN" it writes after reasoning; unparsed dropped): ECE **rises**
0.232 to 0.342 (**+0.073 [+0.009, +0.130]**) - i.e. internal calibration improves while *spoken* confidence
does not. Trust the logits, not the self-report.

**Takeaway:** internal calibration error roughly **quartered** (no "confidence tax" in selective
prediction), at a small accuracy cost. Replicated on MedGemma-4B (ECE 0.304 to 0.126). Full method, CIs,
and the matched-accuracy decomposition: full method, CIs, and code: https://github.com/Mr-Neutr0n/confidence-tax-study

## Usage
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-4B-Instruct-2507"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "<this-adapter-path>")  # load the LoRA adapter
```

## License & citation
- **License:** Apache-2.0 (base + data are Apache-2.0).
- Code and analysis: https://github.com/Mr-Neutr0n/confidence-tax-study