---
license: apache-2.0
base_model: google/gemma-4-E4B-it
tags:
  - gguf
  - clinical-documentation
  - home-health
  - caredraft
---

# CareDraft E4B (fine-tuned)

A LoRA fine-tune of **Gemma 4 E4B QAT** (google/gemma-4-E4B-it), trained by the
CareDraft team to turn rough, speech-to-text home-health dictation into
polished draft clinical documentation across physical therapy,
occupational therapy, speech-language pathology, and skilled nursing.

**Training data is 100% synthetic** — no real patient information, transcripts,
or notes were used or are reproducible from this model. All training pairs
were generated and mechanically validated (numeric-fact preservation, denial/
negation integrity, no unsupported claims) before training.

- **Format:** GGUF, Q4_K_M quantization
- **Size:** ~5.30 GB
- **SHA-256:** `9e03f6af0b55fcb88358624055fa622b64a46b92803599b37e1016aa5eb011ca`
- **Base:** Gemma 4 E4B QAT (4-bit)
- **Intended use:** On-device note generation inside the [CareDraft](https://caredraft.app) app. Not intended for standalone medical use — all output requires clinician review before it becomes part of a medical record.

## License

This model is a modified version of Google's Gemma 4 E4B, which is released
under the [Apache License 2.0](https://ai.google.dev/gemma/apache_2). The
fine-tuned weights are distributed under the same license; see [LICENSE](LICENSE).

**Modifications:** CareDraft LLC fine-tuned the instruction-tuned Gemma 4 E4B QAT
checkpoint with LoRA on synthetic home-health dictation, merged the adapter into
the base weights, and converted and quantized the result to GGUF (Q4_K_M).

This model is not made or endorsed by Google.

## Usage

This model is built for CareDraft's local llama.cpp-based inference pipeline
(Gemma chat template, `<end_of_turn>`/`<eos>` stop tokens) and is not
packaged as a general-purpose chat assistant.