Add model card with Tier A eval results + Ollama/llama.cpp usage

README.md

@@ -1,31 +1,97 @@
 ---
+base_model: unsloth/gemma-4-E4B-it
+library_name: gguf
+license: gemma
+pipeline_tag: text-generation
+language:
+  - en
+  - zh
+  - ar
 tags:
-- gguf
-- llama.cpp
-- unsloth
-- vision-language-model
+  - gemma
+  - gemma-4
+  - gguf
+  - quantized
+  - q4-k-m
+  - llama-cpp
+  - ollama
+  - fine-tuned
+  - rag
+  - offlineaid
+  - australian-consumer-safety
+  - anti-scam
+  - disaster-response
+datasets:
+  - helenkwok/offlineaid
 ---
 
-# gemma_4_finetune : GGUF
+# OfflineAid — Gemma 4 E4B fine-tune (Q4_K_M GGUF)
 
-This model was finetuned and converted to GGUF format using [Unsloth](https://github.com/unslothai/unsloth).
+Q4_K_M-quantized GGUF of the [OfflineAid](https://github.com/helenkwok/offlineaid) Stage-1 fine-tune. Drop-in for `llama.cpp` and Ollama.
 
-**Example usage**:
-- For text only LLMs:    `llama-cli -hf helenk/gemma_4_finetune --jinja`
-- For multimodal models: `llama-mtmd-cli -hf helenk/gemma_4_finetune --jinja`
+- **Base:** [`unsloth/gemma-4-E4B-it`](https://huggingface.co/unsloth/gemma-4-E4B-it)
+- **LoRA:** [`helenk/gemma-4-E4B-lora`](https://huggingface.co/helenk/gemma-4-E4B-lora) — Unsloth fine-tune on Kaggle T4
+- **Merged fp16 source:** [`helenk/gemma-4-E4B-finetune`](https://huggingface.co/helenk/gemma-4-E4B-finetune)
+- **Quantize chain:** `peft.merge_and_unload` → `llama.cpp/convert_hf_to_gguf.py` → `llama-quantize Q4_K_M`
+- **File size:** ~5.0 GB
 
-## Available Model files:
-- `gemma-4-e4b-it.Q4_K_M.gguf`
-- `gemma-4-e4b-it.F16-mmproj.gguf`
+## Tier A held-out eval (vs stock + RAG)
 
-## ⚠️ Ollama Note for Vision Models
-**Important:** Ollama currently does not support separate mmproj files for vision models.
+Held-out: 111 rows stratified per-language (37 EN + 37 ZH + 37 AR) from the 1,113-row [`helenkwok/offlineaid`](https://huggingface.co/datasets/helenkwok/offlineaid) corpus, seed=3407. Both models served by Ollama (Q4_K_M). Greedy decoding, explicit "Answer in {language}" directive.
 
-To create an Ollama model from this vision model:
-1. Place the `Modelfile` in the same directory as the finetuned bf16 merged model
-3. Run: `ollama create model_name -f ./Modelfile`
-   (Replace `model_name` with your desired name)
+| Language | Metric        | stock + RAG | ft + RAG | Δ |
+|----------|---------------|------------:|---------:|---:|
+| EN       | ROUGE-L F1    | 0.688       | **0.699** | +0.011 |
+| EN       | Format-OK %   | 91.9%       | **94.6%** | +2.7 pp |
+| ZH       | Format-OK %   | 45.9%       | **62.2%** | **+16.3 pp** |
+| AR       | ROUGE-L F1    | 0.085       | **0.139** | **+63%** |
+| AR       | Format-OK %   | 21.6%       | **54.1%** | **+32.4 pp (2.5×)** |
+| **all**  | ROUGE-L F1    | 0.334       | **0.355** | +0.021 |
+| **all**  | Format-OK %   | 53.2%       | **70.3%** | **+17.0 pp** |
 
-This will create a unified bf16 model that Ollama can use.
-This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
-[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
+The fine-tune's value lives in **multilingual robustness, especially Arabic** (format-OK 21.6% → 54.1%, ROUGE-L +63%). Reproducible via `bash scripts/tier_a_pipeline_eval_only.sh` in the [OfflineAid repo](https://github.com/helenkwok/offlineaid).
+
+## Use with Ollama
+
+```
+ollama pull hf.co/helenk/gemma-4-E4B-finetune-GGUF
+```
+
+Or via a local `Modelfile`:
+
+```
+FROM /path/to/gemma-4-E4B-offlineaid-Q4_K_M.gguf
+RENDERER gemma4
+PARSER gemma4
+PARAMETER num_ctx 32768
+PARAMETER stop "<turn|>"
+PARAMETER temperature 0.0
+```
+
+```
+ollama create offlineaid-e4b -f Modelfile
+ollama run offlineaid-e4b
+```
+
+## Use with llama.cpp
+
+```
+./llama-cli \
+  -m gemma-4-E4B-offlineaid-Q4_K_M.gguf \
+  -p "Answer in Simplified Chinese.\n\nQUESTION: ..." \
+  --temp 0.0 -n 256
+```
+
+## Intended use
+
+Stage 3 of the OfflineAid pipeline — Mac-side pack-builder agent loop. Pixel 7 production deployment uses *stock* Gemma 4 E2B + retrieval, not this fine-tune; see [the project writeup](https://github.com/helenkwok/offlineaid) for the architectural rationale.
+
+## License
+
+Inherits Google's [Gemma Terms of Use](https://ai.google.dev/gemma/terms). Training data ([`helenkwok/offlineaid`](https://huggingface.co/datasets/helenkwok/offlineaid)) is CC-BY-4.0.
+
+## Sibling repos
+
+- Merged fp16 safetensors (~16 GB): [`helenk/gemma-4-E4B-finetune`](https://huggingface.co/helenk/gemma-4-E4B-finetune)
+- LoRA adapter (~50 MB): [`helenk/gemma-4-E4B-lora`](https://huggingface.co/helenk/gemma-4-E4B-lora)
+- Smaller E2B variant: [`helenk/gemma-4-E2B-finetune-GGUF`](https://huggingface.co/helenk/gemma-4-E2B-finetune-GGUF)