HypergraphFormer

Link to paper: https://arxiv.org/abs/2605.18932 LoRA adapters fine-tuning Qwen/Qwen3-4B-Instruct-2507 for hypergraph-based floorplan generation. The repo contains several adapters trained on different dataset sizes.

Checkpoints

Subfolder	Train samples	Step
qwen_hypergraphformer_1000_samples/checkpoint-240	1,000	240
qwen_hypergraphformer_5000_samples/checkpoint-750	5,000	750
qwen_hypergraphformer_10000_samples/checkpoint-1500	10,000	1500
qwen_hypergraphformer_25000_samples/checkpoint-3900	25,000	3900
qwen_hypergraphformer/checkpoint-8700	full	8700

LoRA configuration

• Rank r = 64, lora_alpha = 128, lora_dropout = 0.1
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Task: CAUSAL_LM

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-4B-Instruct-2507"
repo_id = "NikitaKlimenko/HypergraphFormer"
subfolder = "qwen_hypergraphformer_25000_samples/checkpoint-3900"

tok = AutoTokenizer.from_pretrained(repo_id, subfolder=subfolder)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, repo_id, subfolder=subfolder)