kabuizuchi-trading/gdpo-qwen-structured-merged

SFT + GDPO trained Qwen3-4B for Structured Output

This model was trained in two phases:

1. SFT (Supervised Fine-Tuning) for structured output basics
2. GDPO (Group reward-Decoupled Normalization Policy Optimization) for multi-reward RL refinement

This repository contains the full-merged 16-bit weights.

GDPO Training (Phase 2)

GDPO (arxiv 2601.05242, NVIDIA Research) decouples reward normalization across individual rewards, preserving their relative differences.

Reward Functions

#	Reward	Weight	Description
1	Format Compliance	1.0	Approach/Output structure
2	Structured Output Validity	3.0	JSON/XML/YAML/TOML/CSV parsing
3	Output Length	0.3	Appropriate length
4	No Repetition	0.5	Prevents degeneration

Configuration

• Method: GDPO (reward_aggregation=normalize_then_sum)
• scale_rewards: none
• LoRA: r=16, alpha=32
• Learning rate: 5e-07
• Beta: 0.01
• Num generations: 8
• Max steps: 300
• Max completion length: 384
• Dataset size: 2000 prompts (subsampled)
• GPU: NVIDIA A100 80GB (bf16)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "kabuizuchi-trading/gdpo-qwen-structured-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto",
)

Sources & License

• GDPO Paper: https://arxiv.org/abs/2601.05242
• License: Apache-2.0