TenaOS - Gemma 4 E4B Task-Tagged LoRA
TenaOS is a local-first clinical AI operating system for primary-care workflows. This repository hosts the Gemma 4 E4B runtime artifacts used by TenaOS, including the base BF16 GGUF model, multimodal projector, and the task-tagged LoRA adapter trained for TenaOS clinical-informatics workflows.
TenaOS follows a constrained clinical-agent pattern: Gemma proposes, local knowledge bases ground, deterministic middleware validates, and clinicians review before anything is persisted to OpenMRS.
Release Status
This model card describes the current released adapter and merged GGUF artifacts. The documentation, metadata, and model-card charts have been refreshed to match the released weights.
The adapter was trained from real multi-turn production workflow traces with assistant-turn loss masking. Workflow-level validation is still required before making task-by-task performance claims against the base model.
Files
| File | Purpose |
|---|---|
gemma-4-E4B-it-BF16.gguf | Base Gemma 4 E4B BF16 GGUF used by the local llama.cpp runtime |
mmproj-gemma-4-E4B-it-bf16.gguf | Base multimodal projector for audio input |
adapter/adapter_model.safetensors | TenaOS task-tagged LoRA adapter |
adapter/adapter_config.json | LoRA adapter configuration |
adapter/training_metadata.json | Training configuration and runtime summary |
merged_hf/ | Merged BF16 Hugging Face checkpoint |
tenaos-gemma-4-E4B-it-lora-F16.gguf | Merged LoRA F16 GGUF artifact |
mmproj-tenaos-gemma-4-E4B-it-lora-bf16.gguf | Projector packaged with the merged LoRA GGUF |
tenaos-gemma-4-E4B-it-lora-Q4_K_M.gguf | Optional quantized deployment artifact, when present |
tenaos-technical-report.pdf | Technical report |
training_corpus/ | Synthetic SFT corpus used for adapter training |
training_code/ | Training, merge, conversion, and eval helper scripts |
The base BF16 GGUF and projector filenames are preserved for compatibility with the TenaOS bootstrap scripts.
Task Tags
The adapter is trained as a single multi-task adapter routed by explicit task tags:
| Tag | Workflow |
|---|---|
[form] | Natural-language form and workflow building |
[report] | Plain-language report planning |
[scribe] | English text and voice scribing |
[scribe-am] | Amharic text scribing |
[cds] | Clinical decision support |
[edu] | Patient education material generation |
Training Summary
The adapter was trained on curated, task-tagged clinical-informatics workflow traces reconstructed from the TenaOS production stack.
| Field | Value |
|---|---|
| Base model used for training | unsloth/gemma-4-E4B-it |
| Published base lineage | google/gemma-4-E4B-it |
| Training mode | BF16 LoRA, text decoder only |
| Validated traces | 16,005 |
| Train / validation / test | 18,909 / 1,071 / 1,109 |
| Epochs / steps | 3 / 7,086 |
| LoRA rank / alpha / dropout | r=16 / alpha=32 / dropout=0.0 |
| Max sequence length | 24,576 |
| Loss masking | Assistant turns only |
| Chat template | Native Gemma 4 tokenizer template |
| Runtime | 70.5 hours on A100 80GB |
| Final train loss | 0.04123 |
| 4-bit loading | false |
Dataset And SFT Format
The corpus keeps seven workflow families and uses real multi-turn ShareGPT-style conversations reconstructed from production event traces, including system, user, assistant tool-call, and tool-result turns where available. The training script applies the Gemma 4 chat template and masks loss to assistant turns only.
| Task | Train | Validation | Test |
|---|---|---|---|
[form] | 6,001 | 347 | 376 |
[cds] | 3,407 | 197 | 202 |
[edu] | 3,406 | 198 | 210 |
[report] | 3,134 | 156 | 156 |
[scribe-am] | 1,291 | 88 | 74 |
[scribe] English text | 946 | 50 | 54 |
[scribe] voice/audio | 724 | 35 | 37 |
| Total | 18,909 | 1,071 | 1,109 |
The released training corpus is available under training_corpus/.
It is synthetic, teacher-generated, task-tagged training data, not real patient
records and not clinical validation data. The corresponding training and merge
scripts are available under training_code/.
Merge Details
The released merged model uses checkpoint 7086. The LoRA merge applied the text decoder adapter into BF16 base weights. Vision tower and multimodal projector weights remain base-model weights.
| Field | Value |
|---|---|
| Merge schema | tenaos_lora_merge_v1 |
| Adapter directory | lora_training/runs/20260703T061340Z/adapter/checkpoint-7086 |
| Merged dtype | bfloat16 |
| Language layers merged | 294 |
| Vision layers changed | 0 |
Training Charts
Running With llama.cpp
Base model:
hf download beza4588/TenaOS --local-dir ./models
llama-server \
-m ./models/gemma-4-E4B-it-BF16.gguf \
--mmproj ./models/mmproj-gemma-4-E4B-it-bf16.gguf \
--host 0.0.0.0 \
--port 8000 \
-ngl 99 \
--jinja \
--alias gemma-4
Merged LoRA model, when using the merged GGUF artifact:
llama-server \
-m ./models/tenaos-gemma-4-E4B-it-lora-F16.gguf \
--mmproj ./models/mmproj-tenaos-gemma-4-E4B-it-lora-bf16.gguf \
--host 0.0.0.0 \
--port 8000 \
-ngl 99 \
--jinja \
--alias gemma-4
In TenaOS, the Docker image bind-mounts this directory at /models. See
scripts/fetch-models.sh.
Intended Use
This model package is intended for the TenaOS local clinical AI runtime. It is not intended to autonomously diagnose, prescribe, or write directly to a medical record. TenaOS uses allow-listed tools, local WHO/MSF and CIEL knowledge bases, deterministic validation, and clinician review.
Limitations
- The adapter is trained for TenaOS workflow traces and task tags. It should be evaluated in the full TenaOS runtime rather than as a generic chat model.
- Workflow-level metrics such as form recall, report correctness, scribe extraction quality, unsupported clinical recommendation rate, citation quality, and CDS grounding should be measured in the full runtime.
- Clinical output remains draft material until reviewed by a qualified clinician.
License
The Gemma model artifacts inherit the Gemma Terms of Use. TenaOS packaging and application code are released separately under the Apache 2.0 license.