LoRA adapter (rank 16, α 32, dropout 0.05) for google/gemma-4-E2B-it, fine-tuned to read OBD-II sensor windows and explain car problems in plain English with calibrated repair-cost estimates and the specific upsell traps mechanics attach to each fault.

For inference, most users want the GGUFs: MindFreakGamer/gemma-4-E2B-pocket-mechanic-GGUF. Use this adapter only if you want to fuse into a different base, experiment with merging, or further fine-tune.

Quick load

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-E2B-it")
model = PeftModel.from_pretrained(base, "MindFreakGamer/gemma-4-E2B-pocket-mechanic-lora")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-E2B-it")

Training

LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Data: 4,365 (sensor window → diagnostic explanation) pairs distilled from Claude Opus 4.7 via the Anthropic Batch API
1× A10G on Hugging Face Jobs, Unsloth + TRL, 2 epochs, 2h 42m
Eval loss 3.73 → 2.31

Benchmark

Fused + Q8_0 quantized = 81.3% of Claude Opus 4.7 teacher quality (blind A/B judge, n=100). See the GGUF model card for the full benchmark table.

License

Inherits the Gemma license.

Submission to the Hugging Face Build Small Hackathon, Backyard AI track (June 2026). Code: github.com/small-hack-huggingface/obd-intelligence.