gemma2-handy-lora

What is this?

A fine-tuned google/gemma-2-2b-it (2B params) that acts as a post-processing layer for voice-driven coding tools. It sits between your speech-to-text (STT) app and your code.

The problem it solves

Speech-to-text engines output exactly what you said — full of filler, hedging, and loose phrasing ("kind of don't hit the server so much maybe?"). Code commands need to be one short, unambiguous instruction that an app can act on. This model closes that gap.

What it does

Takes the raw transcription and rewrites it into a single canonical engineering instruction in your voice, picking the precise domain term for what you meant:

you say (STT output)	what you meant	model output
"don't hit the server so much"	rate limiting	Rate-limit the API calls.
"make sure this only runs once"	idempotency	Make the webhook handler idempotent.
"load images only when they appear"	lazy loading	Lazy-load images as they enter the viewport.

It classifies input into 11 canonical domains: rate-limit, cron, mock, debounce, circuit-breaker, optimistic, lazy-load, backpressure, idempotent, blue-green, event-sourcing — and falls back to fixing spelling/filler when input fits no domain.

Why a 2B fine-tune instead of a big API model?

• Private — runs 100% locally (Ollama / llama.cpp); your voice never leaves your machine.
• Fast + free — no per-call cost, no network round-trip, ~30 ms on a Mac.
• Predictable — constrained to canonical domains, so output is app-actionable, not a chatty LLM reply.

Files

• gemma2-q4_k_m.gguf — Q4_K_M quantized model (1.7 GB), ready for llama.cpp / Ollama.
• Modelfile — Ollama model definition (ollama create gemma2:handy-lora -f Modelfile).
• runtime_template.txt — prompt template with exemplars; must be used as the system prompt at inference (ablation: 23/25 with exemplars vs 17/25 without).

Training

• Base: google/gemma-2-2b-it (gated, accepts Gemma license)
• Method: QLoRA-style LoRA (r=16, α=32, dropout 0.1) on all linear projections
• Data: 794 synthetic pairs (11 terms) / 108 holdout
• Recipe: 4 epochs (early stop patience 1), lr 1e-4 cosine, fp16, batch 1 × grad-accum 8, max_grad_norm 0.5; best checkpoint selected by eval loss
• GPU: single Modal L4 (~30 min)

Evaluation

25 never-seen real user prompts (verified disjoint from train + holdout): 23/25 canonical-term hits. Known misses: an "optimistic" like-button update and an "event-sourcing" order-history phrasing (answers are close but not canonical).

Usage

Ollama

ollama create gemma2:handy-lora -f Modelfile
curl http://localhost:11434/v1/chat/completions -d '{
  "model": "gemma2:handy-lora",
  "messages": [
    {"role": "system", "content": "<paste runtime_template.txt>"},
    {"role": "user", "content": "don't hit the server so much"}
  ],
  "temperature": 0.2,
  "max_tokens": 20
}'

llama.cpp

llama-cli -m gemma2-q4_k_m.gguf -p "<prompt>" -n 20 --temp 0.2

Repo

Code, dataset generator, training pipeline, and eval harness: https://github.com/moe5445/gemma2-handy-finetune