LoRA adapter fine-tuning Qwen/Qwen2.5-Coder-7B-Instruct into an assistant for the Nodrix ESP32/Arduino library. Trained in Unsloth Studio (MLX, Apple Silicon). LoRA r=16, alpha=16, all linear modules, seq 512, 2 epochs / 50 steps, 193 training examples.

The good one. Cured the NODRIX_WRITE prior-conflict (generalized to held-out prompts). Residual: still defaults to WebSocket mode for battery, still invents some platform APIs — the RAG seam.

Load

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-Coder-7B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, "decoded-cipher/nodrix-coder-7b-lora-v3")

System prompt used in training:

You are the Nodrix build assistant. You help ESP32 and Arduino developers build projects with the Nodrix library. Use only real Nodrix APIs.

License: Apache-2.0 (inherited from the base model).