zhenchonghu/dsaa6000q-a3-backward

DSAA6000Q Backward LoRA Adapter

This repository contains a LoRA adapter produced for DSAA6000Q Assignment 3 using the self-alignment pipeline from Self Alignment with Instruction Backtranslation.

Summary

Backward model p(instruction | response) trained on OpenAssistant Guanaco.

Base model

• Qwen/Qwen3-1.7B

Training data

• Dataset: timdettmers/openassistant-guanaco split train
• Max train samples: 2000
• Max eval samples: 200

Notes

• This is a PEFT LoRA adapter, not a fully merged checkpoint.
• The model was trained with a standalone Kaggle-compatible notebook.