ravikadam/running-coach-gemma3-1b-GGUF

🏃 Running Coach — Gemma 3 1B (GGUF, Q4KM)

A 1B-parameter running coach fine-tuned from Google Gemma 3 1B to answer running questions — technique, training, injuries, recovery, gear, motivation — and small enough to run fully offline on a phone or laptop (~769 MB, 4-bit Q4KM).

⚠️ Educational demo — NOT medical advice. This model can be wrong. It is not a doctor, physiotherapist, or certified coach. If you have pain, a possible injury, or any medical symptom, stop and consult a qualified professional. Do not use it for health decisions.

Use with Ollama

# Modelfile
printf 'FROM ./running-coach-gemma3-1b-Q4_K_M.gguf\nTEMPLATE """<start_of_turn>user\n{{ .Prompt }}<end_of_turn>\n<start_of_turn>model\n"""\nPARAMETER stop "<end_of_turn>"\n' > Modelfile
ollama create running-coach -f Modelfile
ollama run running-coach "My knee hurts below the kneecap after running, what should I do?"

Also works with llama.cpp, LM Studio, and other GGUF runtimes. For Google AI Edge Gallery you need the LiteRT .task build (see the project repo).

How it was built

• —Dataset: ~1,200 instruction/response pairs synthesized from a running guide, semantically de-duplicated, with safety/identity seeds.
• —Fine-tune: QLoRA, 2 epochs, trained locally with Unsloth Desktop (Apple MLX). Train loss 2.33 → 1.05.
• —Eval (DeepEval, judge gpt-5.4-mini): coach identity 0%→100%, coaching quality 35%→81%, answer relevancy 96%→100%. Known weak spots: off-topic redirects ~50%, and medical-emergency escalation regressed to 50% — hence the disclaimer above.

Attribution

Base model Gemma 3 © Google, under the Gemma Terms of Use. Coaching content paraphrased from a third-party running guide; this project is educational and not affiliated with or endorsed by its authors. Source text and training dataset are not redistributed.

Project & code: https://github.com/ravikadam/running-coach-gemma3-1b