eoinedge/arduino-edgeai-qwen-combined-0.5b-lora

Arduino + Edge AI Combined Qwen 0.5B LoRA Adapter

This adapter fine-tunes Qwen/Qwen2.5-Coder-0.5B-Instruct on a combined Arduino + Edge AI corpus to provide a single offline-capable model for both Arduino documentation guidance and Edge AI tooling knowledge.

Model Details

• —Model type: LoRA adapter
• —Base model: Qwen/Qwen2.5-Coder-0.5B-Instruct
• —Adapter size: 0.5B
• —Training corpora: subset/subset_arduino + subset/subset_edgeai
• —Task: text generation / coding assistance
• —Library: peft
• —Pipeline: text-generation

Intended Use

This adapter is intended for users who need offline or local inference with both Arduino documentation knowledge and Edge AI reference material in one model.

Training Details

• —Training strategy: LoRA fine-tuning on combined subset corpora
• —Hyperparameters: 3 epochs, batch size 2, gradient accumulation 4, learning rate 1e-4
• —Data processed: combined subset of Arduino docs and Mintlify Edge AI docs
• —Output directory: adapter-combined-subset-0.5b

Evaluation

QA Evaluation

• —Dataset: 10 fixed Arduino-domain questions
• —Base model avg response time: 6.41s
• —Adapter avg response time: 6.57s
• —Base keyword hits avg: 7.0
• —Adapter keyword hits avg: 5.5
• —Base code snippet presence: 9/10
• —Adapter code snippet presence: 1/10
• —Result file: eval_combined_0.5b_results.csv

Perplexity Evaluation

• —Test corpus: subset/subset_edgeai (121 docs)
• —Base mean PPL: 9.46
• —Adapter mean PPL: 6.65
• —Mean delta: -2.81
• —Adapter wins: 117 / 121 files
• —Result file: ppl_combined_0.5b_subset_edgeai_results.csv

Notes

• —Lower perplexity indicates the adapter is more confident on Edge AI domain text.
• —The QA evaluation is a small fixed question set, and the perplexity results are the deeper numeric signal for domain adaptation.

References

• —Offline SLMs for Edge AI Development — Part 1: Qwen LoRA Adapter Fine-Tuned on Edge Impulse Docs
• —Offline SLMs for Edge AI Development — Part 2: RAG as an Enhancement for Fine-Tuned Models with FAISS
• —Offline SLMs for Edge AI Development — Part 3: Agentic Coding with an Arduino Fine-Tuned Adapter via llama.cpp and OpenCode

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from peft import PeftModel

base = "Qwen/Qwen2.5-Coder-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, "eoinedge/arduino-edgeai-qwen-combined-0.5b-lora")
model = model.merge_and_unload()
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device="cpu")
print(pipe("How do I read a DHT11 sensor in Arduino?", max_new_tokens=200)[0]["generated_text"])

License

This adapter is derived from the base model and follows the licensing terms of the base model and related adapter resources.