webAI-Official/math-merged-adapter

Combined SFT + GRPO LoRA Adapter for Qwen3-4B

This adapter combines two LoRA training stages into a single adapter:

1. SFT (Supervised Fine-Tuning) on Qwen/Qwen3-4B
2. GRPO (Group Relative Policy Optimization) on the SFT model

The two rank-32 adapters were merged into a single rank-64 adapter (lossless). Apply directly to Qwen/Qwen3-4B — no intermediate merged model needed.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B")
model = PeftModel.from_pretrained(base_model, "abdul-hannan/qwen3-math-grpo")
tokenizer = AutoTokenizer.from_pretrained("abdul-hannan/qwen3-math-grpo")

Training Details

• Base model: Qwen/Qwen3-4B
• LoRA rank: 64 (combined from two rank-32 adapters)
• LoRA alpha: 128
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• PEFT version: 0.18.1

Contact

Syed Abdul Hannan

Framework versions

• PEFT 0.18.1