qwen3-1.7b-backward-lora

Qwen3-1.7B Backward LoRA (p(x|y))

This adapter turns Qwen3-1.7B into a backward model that generates an instruction from a given response. It is used in the self-alignment pipeline to create synthetic instruction-response pairs from unlabeled answers.

Training data

• OpenAssistant Guanaco train set (first-turn only)
• Each example parsed into (instruction, output) and flipped to (output -> instruction)

Objective

Causal LM with loss applied only on the instruction tokens.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-1.7B"
adapter = "EvaanLi/qwen3-1.7b-backward-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)

Notes

This adapter is for instruction generation only and is not intended for direct instruction-following.