Base: unsloth/gemma-4-E4B-it
LoRA: helenk/gemma-4-E4B-lora — Unsloth fine-tune on Kaggle T4
Merged fp16 source: 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

Tier A held-out eval (vs stock + RAG)

Held-out: 111 rows stratified per-language (37 EN + 37 ZH + 37 AR) from the 1,113-row helenkwok/offlineaid corpus, seed=3407. Both models served by Ollama (Q4_K_M). Greedy decoding, explicit "Answer in {language}" directive.

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

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.

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 for the architectural rationale.

License

Inherits Google's Gemma Terms of Use. Training data (helenkwok/offlineaid) is CC-BY-4.0.

Sibling repos

Merged fp16 safetensors (~16 GB): helenk/gemma-4-E4B-finetune
LoRA adapter (~50 MB): helenk/gemma-4-E4B-lora
Smaller E2B variant: helenk/gemma-4-E2B-finetune-GGUF