qwen3-1.7b-final-lora

Qwen3-1.7B Final LoRA (instruction-following)

This adapter is the final model in the self-alignment pipeline. It is trained on self-curated instruction-response pairs produced by the backward model and scored by the base model.

Training data

• Generated (instruction, response) pairs from LIMA single-turn answers
• Self-curated with the Table 19 prompt from the paper (score >= 4)

Objective

Causal LM with loss applied only on the response tokens.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-1.7B"
adapter = "EvaanLi/qwen3-1.7b-final-lora"

tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)