Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Long Xia, Dawei Yin and Chao Huang*. (*Correspondence )
Data Intelligence Lab@University of Hong Kong, Baidu Inc.
This repository hosts the code, data and model weight of HiGPT.
- 🚀 [2024.05] 🎯🎯📢📢 Our 🌟HiGPT🌟 has been accepted by KDD'2024 Research Track (20% acceptance rate)! Congrats to all HiGPT team! 🎉🎉🎉
🎯🎯📢📢 We have made significant updates to the models used in our HiGPT on 🤗 Huggingface. We highly recommend referring to the table below for further details:
| 🤗 Huggingface Address | 🎯 Description |
|---|---|
| https://huggingface.co/Jiabin99/In-Context-HGT | The trained in-context heterogeneous graph tokenizer using our lightweight text-graph contrastive alignment. |
| https://huggingface.co/Jiabin99/HiGPT | It's the checkpoint of our HiGPT based on Vicuna-7B-v1.5 tuned on 60 shots IMDB graph instruction data. |
- [2024.02.24]🔥🔥Release our utilized Instruction data.
- [2024.02.24]🔥🔥Release checkpoints of our HiGPT and pre-trained graph encoder.
- [2024.02.24] 🚀🚀 Release the code of HiGPT.
- Supporting lightning training
- Releasing the Chinese version of the explanation
- Releasing the full paper of our HiGPT
- Exploring the potential of our HiGPT for more graph learning tasks.
- ...
we present the HiGPT framework that aligns LLMs with heterogeneous graph structural knowledge with a heterogeneous graph instruction tuning paradigm.
- In-Context Heterogeneous Graph Tokenizer. To achieve adaptability in a wide range of heterogeneous graph sce- narios with varying node and edge types, we introduce the in- context heterogeneous graph tokenizer. This tokenizer captures the diverse semantic relationships found in different heterogeneous graphs, providing a unified approach. To optimize performance and integrate the tokenizer seamlessly into the HiGPT framework, we employ pre-training with a lightweight text-graph contrastive alignment paradigm. For pretraining details, please refer to [./HG_grounding].
- Heterogeneous Graph Instruction-Tuning. We intro- duce a novel heterogeneous graph instruction-tuning framework that integrates inter-type and intra-type token matching tasks to fine-tune large language models (LLMs). Our framework specifically targets the enhancement of LLMs’ understanding of both hetero- geneous relation awareness and homogeneous relation awareness. By utilizing these tasks, our aim is to bolster the LLMs’ capabilities in the following areas: (i) distinguishing between different types of graph tokens, (ii) comprehending intricate relationships within heterogeneous graphs, (iii) preserving the distinctive attributes of entities within homogeneous graphs, and (iv) effectively harnessing diverse graph instructions during the training process. Please refer to Getting Started to explore more.
- Mixture-of-Thought Augmentation. Our approach introduces a novel mechanism for augmenting graph instructions, emphasizing the use of Mixture-of-Thought (MoT) combined with various prompting techniques. This integration enables us to gen- erate a diverse and comprehensive set of informative task-specific instructions. By seamlessly incorporating these augmented graphinstructions into our framework, we anticipate that our model en- hancement will effectively address the challenge of data sparsity. For prompting examples, please refer to [./mot_prompting].
For more technical details, kindly refer to the paper and the project website of our Graph.
1. Environment Preparation [Back to Top]
Please first clone the repo and install the required environment, which can be done by running the following commands:
conda create -n higpt python=3.8
conda activate higpt
# Torch with CUDA 11.7
pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu117
# To support vicuna base model
pip3 install "fschat[model_worker,webui]"
# To install pyg and pyg-relevant packages
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.13.0+cu117.html
# Clone our HiGPT
git clone https://github.com/HKUDS/HiGPT.git
cd HiGPT
# Install required libraries
pip install -r requirements.txt2. Data Preparation [Back to Top]
🎯 (Temporarily recommended from 2024.03.25) Due to the large size of instruction data in the stage 1, we have not upload the complete stage 1 instruction data to the archive. We have an alternative way to download the complete version of stage 1 instruction data: https://pan.baidu.com/s/1rtdxzhRRcEUEpXoWKml36A?pwd=id6u. And we will upload the complete version to the archive as soon as possible. But other data used by HiGPT could be accessed by archive.
(Temporarily deprecated from 2024.03.25) The tuning data of our HiGPT consists of two parts, i.e., heterogeneous graph corpus (stage 1) and heterogeneity-aware graph instruction (stage 2). You can cd hi_datasets and run sh get_stage1_data.sh to download the data in stage 1:
cd /path/to/HiGPT/hi_datasets
wget https://archive.org/download/higpt_stage1/matching_instruction.tar.gz
tar -xzvf matching_instruction.tar.gz
rm -f matching_instruction.tar.gz(Always recommended) Also, you can run sh get_stage1_data.sh to download the data in stage 2:
cd /path/to/HiGPT/hi_datasets
mkdir stage2_data
cd stage2_data
wget https://archive.org/download/higpt_stage2/instruct_ds_dblp.tar.gz
wget https://archive.org/download/higpt_stage2/processed_dblp.tar.gz
wget https://archive.org/download/higpt_stage2/instruct_ds_imdb.tar.gz
wget https://archive.org/download/higpt_stage2/processed_imdb.tar.gz
wget https://archive.org/download/higpt_stage2/instruct_ds_acm.tar.gz
wget https://archive.org/download/higpt_stage2/processed_acm.tar.gz
mkdir DBLP
mkdir IMDB
mkdir acm
tar -xzvf instruct_ds_dblp.tar.gz -C DBLP
tar -xzvf processed_dblp.tar.gz -C DBLP
tar -xzvf instruct_ds_imdb.tar.gz -C IMDB
tar -xzvf processed_imdb.tar.gz -C IMDB
tar -xzvf instruct_ds_acm.tar.gz -C acm
tar -xzvf processed_acm.tar.gz -C acm
rm -f instruct_ds_dblp.tar.gz
rm -f processed_dblp.tar.gz
rm -f instruct_ds_imdb.tar.gz
rm -f processed_imdb.tar.gz
rm -f instruct_ds_acm.tar.gz
rm -f processed_acm.tar.gz3. Training HiGPT [Back to Top]
HiGPT tuning paradigm consists of two stages: (1) instruction tuning with heterogeneous graph corpus; (2) heterogeneity-aware fine-tuning.
3.0. Offline Heterogeneous Graph Tokenizing [Back to Top]
Since the graph tokenizer does not update parameters during the two training processes, we use the Offline Heterogeneous Graph Tokenizing method to preprocess the instruction data in order to accelerate the speed of model training. The data downloaded in Data Preparation has been processed with a pre-trained graph tokenizer. If you need to process with your own graph tokenizer, you can refer to the following commands:
- processing a single instruction file and its corresponding graph data (run_graph_tokenizer_single.sh):
cd /path/to/HiGPT
ann_path=./hi_datasets/stage2_data/IMDB/instruct_ds_imdb/ann/IMDB_test_std_0_1000.json
data_type=imdb
graph_path=./hi_datasets/stage2_data/IMDB/instruct_ds_imdb/graph_data/test
python run_offline_hgt_tokenizer_single.py --ann_path ${ann_path} \
--data_type ${data_type} \
--graph_path ${graph_path}- processing instruction files within a directory and their corresponding graph data (run_graph_tokenizer.sh):
cd /path/to/HiGPT
# offline tokenizing for stage1 matching instruction
data_type=imdb
graph_root=./hi_datasets/matching_instruction
dsname_list=(instruct_ds_matching_movie instruct_ds_node_matching)
pretrained_hgt=./MetaHGT_imdb_dblp_epoch5
for dsname in "${dsname_list[@]}"
do
python run_offline_hgt_tokenizer.py --dsname ${dsname} \
--data_type ${data_type} \
--pretrained_gnn_path ${pretrained_hgt} \
--graph_root ${graph_root}
done
data_type=dblp
graph_root=./hi_datasets/matching_instruction
dsname_list=(instruct_ds_matching_author instruct_ds_matching_paper instruct_ds_node_matching)
pretrained_hgt=./MetaHGT_imdb_dblp_epoch5
for dsname in "${dsname_list[@]}"
do
python run_offline_hgt_tokenizer.py --dsname ${dsname} \
--data_type ${data_type} \
--pretrained_gnn_path ${pretrained_hgt} \
--graph_root ${graph_root}
done3.1. Preparing Pre-trained Checkpoint [Back to Top]
HiGPT is trained based on following excellent existing models. Please follow the instructions to prepare the checkpoints.
Vicuna: Prepare our base model Vicuna, which is an instruction-tuned chatbot and base model in our implementation. Please download its weights here. We generally utilize v1.1 and v1.5 model with 7B parameters.Pretrained Graph Tokenizer: is used to encode heterogeneous graph structures. We employ text-graph grounding approach to obtain the pre-trained heterogeneous graph transformer model, which you could download by heterogeneous graph transformer and put it at [./HiGPT]. We also provide source codes for text-graph grounding at [./HG_grounding] for your reference.
3.2. Instruction Tuning with Heterogeneous Graph Corpus [Back to Top]
You could start the first stage tuning by filling blanks at higpt_stage_1.sh. There is an example as below:
# to fill in the following path to run the first stage of our HiGPT!
#!/bin/bash
cd /path/to/HiGPT
data_path=instruct_ds_matching_author,instruct_ds_matching_movie,instruct_ds_matching_paper,instruct_ds_node_matching,instruct_ds_node_matching_imdb
graph_root=./hi_datasets/matching_instruction
output_dir=./checkpoints/higpt-metahgt-stage1-7b-dblp-imdb-epoch1
base_model=/path/to/vicuna-7b-v1.5-16k
wandb offline
python3.8 -m torch.distributed.run --nnodes=1 --nproc_per_node=4 --master_port=20001 \
higpt/train/train_hete_nopl.py \
--model_name_or_path ${base_model} \
--version v1 \
--data_path ${data_path} \
--graph_content /root/paddlejob/workspace/env_run/llm/GraphChat/playground/data/arxiv_ti_ab.json \
--graph_root ${graph_root} \
--graph_tower ${graph_tower} \
--tune_graph_mlp_adapter True \
--graph_select_layer -2 \
--use_graph_start_end True \
--bf16 False \
--output_dir ${output_dir} \
--num_train_epochs 1 \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 1 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 2400 \
--save_total_limit 1 \
--learning_rate 2e-5 \
--weight_decay 0.0001 \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 2048 \
--gradient_checkpointing True \
--lazy_preprocess True \
--report_to wandb \
--hetero_key_path /root/paddlejob/workspace/env_run/output/sample_instruct_ds/ann/hetero_key_order.json3.3. Extract the Trained Projector [Back to Top]
We could extract the trained projector in the stage 1 by filling blanks at extract_projector.sh. There is an example as below:
# to fill in the following path to extract projector for the first tuning stage!
#!/bin/bash
cd /path/to/HiGPT
stage1_model=./checkpoints/higpt-metahgt-stage1-7b-dblp-imdb-epoch1
graph_projector=./checkpoints/higpt_stage1_projector_metahgt_dblp_imdb_epoch1/higpt-metahgt-stage1-7b.bin
python3.8 ./scripts/extract_graph_projector.py \
--model_name_or_path ${stage1_model} \
--output ${graph_projector}3.4. Heterogeneity-aware Fine-tuning [Back to Top]
You could start the second stage tuning based on different number of shots (e.g., 1, 3, 5, 10, 20, 40, 60) by filling blanks at higpt_stage_2.sh. There is an example as below:
# to fill in the following path to run the second stage of our HiGPT!
#!/bin/bash
cd /path/to/HiGPT
base_model=/path/to/vicuna-7b-v1.5-16k
graph_root=./hi_datasets/stage2_data/IMDB
graph_projector=./checkpoints/higpt_stage1_projector_metahgt_dblp_imdb_epoch1/higpt-metahgt-stage1-7b.bin
num_epochs=15
num_shot_list=(1 3 5 10 20 40 60)
for num_shot in "${num_shot_list[@]}"
do
python3.8 -m torch.distributed.run --nnodes=1 --nproc_per_node=4 --master_port=20010 higpt/train/train_hete_nopl.py \
--model_name_or_path ${base_model} \
--version v1 \
--data_path instruct_ds_imdb \
--graph_content /root/paddlejob/workspace/env_run/llm/GraphChat/playground/data/arxiv_ti_ab.json \
--graph_root ${graph_root} \
--graph_tower MetaHGT_imdb_dblp_epoch5 \
--pretrain_graph_mlp_adapter ${graph_projector} \
--tune_graph_mlp_adapter True \
--graph_select_layer -2 \
--use_graph_start_end True \
--bf16 False \
--output_dir ./checkpoints/higpt-stage2-imdb-metahgt-epoch${num_epochs}-mixcot-true-${num_shot} \
--num_train_epochs ${num_epochs} \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 1 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 2400 \
--save_total_limit 1 \
--learning_rate 2e-5 \
--weight_decay 0.0001 \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 2048 \
--gradient_checkpointing True \
--lazy_preprocess True \
--report_to wandb \
--hetero_key_path /root/paddlejob/workspace/env_run/output/sample_instruct_ds/ann/hetero_key_order.json \
--num_shot ${num_shot}
done
5. Evaluating HiGPT [Back to Top]
5.1. Preparing Checkpoints [Back to Top]
Checkpoints: You could try to evaluate HiGPT by using your own model or our released checkpoints.
5.2. Running Evaluation [Back to Top]
- You could evaluate our HiGPT by filling blanks at higpt_info_imdb_cot.sh. There is an example as below:
# to fill in the following path to extract projector for the second tuning stage!
#!/bin/bash
cd /path/to/HiGPT
output_model=./checkpoints
datapath=./hi_datasets/stage2_data/IMDB
res_path=./output_res_imdb
num_epochs=15
num_shot_list=(1 3 5 10 20 40 60)
for num_shot in "${num_shot_list[@]}"
do
for ((cot_case=0; cot_case<=0; cot_case++))
do
python3.8 ./higpt/eval/run_higpt.py --model-name ${output_model}/higpt-stage2-imdb-metahgt-epoch${num_epochs}-mixcot-true-${num_shot} --prompting_file ${datapath}/instruct_ds_imdb/ann_processed_MetaHGT_imdb_dblp_epoch5/cot_test/IMDB_test_std_0_1000_cot_${cot_case}.json --graph_root ${datapath} --output_res_path ${res_path}/imdb_test_res_epoch_${num_epochs}_std_${num_shot}_shot_cot_${cot_case} --start_id 0 --end_id 1000 --num_gpus 4
done
done- You could evaluate our HiGPT using the Graph In-Context Learning (Graph ICL) by filling blanks at hetegpt_info_imdb_cot_incontext.sh. There is an example as below:
#!/bin/bash
cd /path/to/HiGPT
output_model=./checkpoints
datapath=./hi_datasets/stage2_data/IMDB
res_path=./output_res_imdb
num_epochs=15
num_shot_list=(1 3 5 10 20 40 60)
cot_case_list=(0)
incontext_dir=in_context_1_shot
for num_shot in "${num_shot_list[@]}"
do
for cot_case in "${cot_case_list[@]}"
do
python3.8 ./higpt/eval/run_higpt_incontext.py --model-name ${output_model}/hetegpt-stage2-imdb-metahgt-epoch${num_epochs}-mixcot-true-${num_shot} --prompting_file ${datapath}/instruct_ds_imdb/ann_processed_MetaHGT_imdb_dblp_epoch5/${incontext_dir}/IMDB_test_std_0_1000_cot_${cot_case}_in_context.json --graph_root ${datapath} --output_res_path ${res_path}_${incontext_dir}/imdb_test_res_epoch_${num_epochs}_std_${num_shot}_shot_cot_${cot_case} --start_id 0 --end_id 1000 --num_gpus 4
done
done
For any questions or feedback, feel free to contact Jiabin Tang.
If you find HiGPT useful in your research or applications, please kindly cite:
@articles{tang2024higpt,
title={HiGPT: Heterogeneous Graph Language Model},
author={Jiabin Tang and Yuhao Yang and Wei Wei and Lei Shi and Long Xia and Dawei Yin and Chao Huang},
year={2024},
eprint={2402.16024},
archivePrefix={arXiv},
primaryClass={cs.CL}
}You may refer to related work that serves as foundations for our framework and code repository, Vicuna, LLaVa, GraphGPT, We also partially draw inspirations from MiniGPT-4. The design of our website and README.md was inspired by NExT-GPT. Thanks for their wonderful works.
