Aurumdev95/gemma4-31b-cessna

Gemma 4 31B — Cessna 172S POH assistant (LoRA)

QLoRA adapter fine-tuning Gemma 4 31B to answer questions about the Cessna 172S avionics equipment, from a synthetic Q&A dataset grounded in the aircraft's Pilot's Operating Handbook (POH) Supplements volume.

• —Base model: unsloth/gemma-4-31B-it-unsloth-bnb-4bit
• —Method: QLoRA (4-bit), LoRA rank 32, language layers only, trained with Unsloth on an NVIDIA DGX Spark (GB10).
• —Data: ~850 grounded Q&A pairs (equipment-disambiguated, de-noised), best-by-validation checkpoint.

Load

from unsloth import FastModel
model, tok = FastModel.from_pretrained("Aurumdev95/gemma4-31b-cessna", load_in_4bit=True, device_map={"": 0})
FastModel.for_inference(model)

Scope & limitations

• —*Coverage is the POH Supplements only (avionics: KMA 28, KX 155A, KAP 140, KLN 94, KR 87, etc.). Core-airframe topics (stall warning, electrical/ammeter, performance, limitations) are not in the source data — answers there rely on the base model's general knowledge and are not POH-verified*.
• —A few counterintuitive facts (e.g. the KAP 140 being rate-/turn-coordinator-driven rather than attitude-gyro-driven) may still be answered from the base model's prior.
• —Not for operational flight use. Always consult the actual POH/AFM.