emorywang/opsd-qwen3-1.7b-singlegpu

OPSD-Qwen3-1.7B (Single-GPU H20 Reproduction)

On-Policy Self-Distillation (OPSD) LoRA adapter for Qwen3-1.7B, reproduced on a single NVIDIA H20 96GB.

Training Config

• Base model: Qwen/Qwen3-1.7B
• Method: OPSD (fixed teacher, LoRA student)
• LoRA: r=64, alpha=128, target=[q,k,v,o,gate,up,down]_proj
• Trainer: TRL 0.26 GOLD trainer
• Steps: 125 (effective batch=8)
• Hardware: 1×H20 96GB, SDPA attention
• Dataset: siyanzhao/Openthoughts_math_30k_opsd (29,434 examples)

Files

• adapter_*.json/safetensors — Final LoRA adapter (step 125)
• checkpoint-100/ — Full training state at step 100 (for paper comparison)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "emorywang/opsd-qwen3-1.7b-singlegpu")
tokenizer = AutoTokenizer.from_pretrained("emorywang/opsd-qwen3-1.7b-singlegpu")