---
license: apache-2.0
base_model: google/gemma-4-E4B-it
language:
- en
- fr
- rw
library_name: gguf
pipeline_tag: text-generation
tags:
- clinical-nlp
- information-extraction
- structured-output
- json
- gguf
- lora
- gemma4
---

# medx-gemma4-e4b

A fine-tune of [Gemma 4 E4B-it](https://huggingface.co/google/gemma-4-E4B-it) that reads one clinical note and returns the
facts as a single flat JSON object. The schema goes in the system prompt: the default schema has 44 fields
(demographics, vitals, social history, symptoms, history, medications, labs, assessment), and a deployer can rename
fields, take a subset, add typed custom fields or ask for nulls without retraining. Notes may be in English, French,
Kinyarwanda or a mix. The model answers directly, with thinking off.

**Research model, not a medical device.** It was trained on model-generated notes whose labels no clinician has
reviewed. Do not use its output for clinical decisions without human verification.

## Files

| File | What it is |
|---|---|
| `medx-gemma4-e4b.Q8_0.gguf` | The shipped model, merged and quantized to Q8_0 for llama.cpp (8.0 GB) |
| `medx-gemma4-e4b.BF16-mmproj.gguf` | Vision projector exported with it; not needed for text-only extraction |
| `adapters/run2_brand/` | LoRA adapter (rank 32, alpha 32) for `google/gemma-4-E4B-it`: stage 1, 30,000 English rows with brand-name drug mentions |
| `adapters/topup/` | LoRA adapter (rank 32, alpha 32) for the **merged stage-1 model**: stage 2, 7,980 rows of French, Kinyarwanda, long English dictations and fresh English notes |

To rebuild the model in Transformers: load `google/gemma-4-E4B-it`, apply `adapters/run2_brand`, merge it, then apply
`adapters/topup` and merge again. The GGUF is the result of exactly that.

## Usage (llama.cpp)

```bash
llama-server -m medx-gemma4-e4b.Q8_0.gguf -c 8192 --jinja \
  --chat-template-kwargs '{"enable_thinking": false}' --temp 0
```

Send the schema as the system message and the note as the user message. Thinking must stay off: the model was
trained without reasoning and the chat template otherwise opens a thought channel. Use greedy decoding.

## Evaluation

Micro-F1 over gold facts on six held-out sets (no note or schema in them was trained on). Real is 150 de-identified
MTSamples transcriptions; the other five sets are model-generated. Labels come from model annotators with model
adjudication.

| Set | Notes | Micro-F1 | Precision | Recall |
|---|---|---|---|---|
| Synthetic (English, default schema) | 987 | 0.945 | 0.947 | 0.943 |
| Hard cases (English) | 250 | 0.939 | 0.942 | 0.936 |
| Real transcriptions (English) | 150 | 0.734 | 0.768 | 0.702 |
| Custom held-out schemas (English) | 1,000 | 0.950 | 0.948 | 0.952 |
| French, held-out schemas | 300 | 0.928 | 0.927 | 0.929 |
| Kinyarwanda, held-out schemas | 300 | 0.910 | 0.906 | 0.915 |

Every output on these sets was valid JSON. Scores are from greedy decoding of the bf16 weights; the Q8_0 GGUF was not
scored separately.

## Training

Unsloth 16-bit LoRA on all attention and MLP projections of the language layers, answer-only loss, 8-bit AdamW,
cosine schedule, effective batch 32, context 4,096, one NVIDIA GB10 (DGX Spark). Stage 1: one epoch on 30,000 rows,
learning rate 1e-4, with half of the generic drug mentions in each note rewritten as brand names. Stage 2: one epoch
on 7,980 rows at 5e-5, starting from the merged stage-1 model.

Every training note was generated from a seeded specification and kept only if each gold fact had a verbatim quote in
the note and every number was reproducible from it; a second model extracted the facts blind and a third adjudicated
disagreements.

## Limitations

- No clinician has reviewed the training or evaluation labels.
- Real-note accuracy (0.734) is far below accuracy on generated notes; the real set has only 150 English notes from
  one public source, and no real French or Kinyarwanda notes have been scored.
- The model can map brand names to the wrong generic drug.
- A larger successor (Qwen3.6-35B-A3B fine-tune) scores higher on every set above.