Model type: PEFT LoRA adapter
Base model: Qwen/Qwen2.5-Coder-1.5B-Instruct
Adapter path: adapter-edgeai-1.5b-full
Target use case: Edge Impulse documentation, Edge AI developer support, and deployment/helpdesk-style answers

Training

Trained on curated Edge AI and Edge Impulse documentation.
Output adapter directory: adapter-edgeai-1.5b-full
Adapter format: safetensors
Training checkpoints saved at steps 1700 and 1764

Evaluation

Perplexity on Edge AI subset:
- base model: 7.30
- adapter: 6.12
- relative improvement: 16.3%
QA coverage on Edge AI question set:
- average keyword match: 5.20 for both base and adapter

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = 'Qwen/Qwen2.5-Coder-1.5B-Instruct'
tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(base_model, trust_remote_code=True)
adapter = PeftModel.from_pretrained(base, 'eoinedge/edgeai-qwen2.5coder-1.5b-lora')

# generate with the adapter
prompt = 'Explain how Edge Impulse can deploy a model to Arduino.'
input_ids = tokenizer(prompt, return_tensors='pt').input_ids
output = adapter.generate(input_ids, max_new_tokens=150)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Files included

adapter-edgeai-1.5b-full/adapter_model.safetensors
adapter-edgeai-1.5b-full/adapter_config.json
adapter-edgeai-1.5b-full/tokenizer.json
adapter-edgeai-1.5b-full/tokenizer_config.json
adapter-edgeai-1.5b-full/chat_template.jinja
eval_edgeai_1.5b_full_edgeai_questions_results.csv
ppl_edgeai_1.5b_full_subset_results.csv

Notes

If you want to load the adapter locally, use PeftModel.from_pretrained with the base model and this adapter repo.
The card will be updated with more detailed dataset provenance and fine-tuning configuration once available.