from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-3B-Instruct"
adapter = "tomyimkc/sophia-agi-lora-v1"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter)

Usage

Train locally:

bash

pip install -r requirements-lora.txt
python tools/prepare_lora_dataset.py
python tools/train_lora.py --4bit --epochs 3

Evaluate:

python tools/eval_local_model.py --adapter training/lora/checkpoints/sophia-v1 --with-gate

Ollama:

ollama create sophia-7b -f models/ollama/Modelfile

Always pair with runtime gate

sophia_gate_check (MCP) or agent/gate.py — weights alone do not guarantee trap safety.