zmli/G-Substrate-Qwen3-VL-2B

G-Substrate (Qwen3-VL-2B)

Multi-task fine-tuned model from the paper "Graph is a Substrate Across Data Modalities" (ICML 2026).

Model Description

This model is fine-tuned from Qwen3-VL-2B-Instruct using the G-Substrate framework, which organizes learning around shared graph structures across heterogeneous modalities and tasks.

Training Details

• Base model: Qwen3-VL-2B-Instruct
• Training method: Full fine-tuning (SFT) with multi-task learning
• Training data: Scene graphs, event graphs, molecular graphs, graph algorithmic tasks, and interleaved role-based data
• Epochs: 2
• Learning rate: 8e-6 (cosine schedule, warmup 10%)
• Batch size: 1 per device, 32 gradient accumulation steps
• GPUs: 2x NVIDIA A100 (DeepSpeed ZeRO-3)

Supported Tasks

Task	Domain	Input	Output
Scene Graph Generation	Visual	Image	Graph triplets
Event Relation Extraction	Text	Document + events	Event relation graph
Molecular Graph Description	Scientific	SMILES + graph	Natural language description
Graph Algorithmic Reasoning	Algorithmic	Graph description	Algorithm answer

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("path/to/G-Substrate-Qwen3-VL-2B", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("path/to/G-Substrate-Qwen3-VL-2B", trust_remote_code=True)

For batch inference with vLLM, see the G-Substrate repository.

Results

CT	CD	SP	BM	BLEU-4	ROUGE-L	PCIs R@50	MA-S F1	MA-T F1	MA-C F1	HiE F1
98.41	96.97	48.59	94.54	51.53	68.47	25.38	52.20	42.68	40.91	25.15

Links

• Code: zmli6/G-Substrate
• Dataset: zmli/G-Substrate-Data

Citation

@inproceedings{li2026gsubstrate,
  title={Graph is a Substrate Across Data Modalities},
  author={Li, Ziming and Wu, Xiaoming and Wang, Zehong and Li, Jiazheng and Tian, Yijun and Bi, Jinhe and Ma, Yunpu and Ye, Yanfang and Zhang, Chuxu},
  booktitle={ICML},
  year={2026}
}