G2S-tkg-forecasting

This is the official code for paper: "G2S: A General-to-Specific Learning Framework for Temporal Knowledge Graph Forecasting with Large Language Models", ACL 2025 Findings.

Develop platform

• 4 A800 (80G) GPUs
• cuda==12.2
• python==3.11
• llama_factory>=0.9.1

Usage

[Optional] Create virtual environment:

conda create -n g2s python=3.11

Install dependencies:

pip install -r requirements.txt

Run unit tests to ensure each module works:

python -m unittest discover tests

Notice: We found that the old version llama3 performs better than current version in our experiments. The main difference is the tokenizer. Please use the provided old version tokenizer llama3_tokenizer/ to reproduce the results.

General Learning Stage

Data preparation:

bash scripts/general_learning/prepare.sh

Training:

bash scripts/general_learning/train-GDELT-WIKI-130k-RID.sh

Specific Learning Stage

Standard setting

Data preparation:

bash scripts/specific_learning/standard/prepare.sh

Run training and evaluation with GID strategy on ICEWS14 dataset:

bash scripts/specific_learning/run-ICEWS14-GID.sh

Zero-shot setting

Data preparation:

bash scripts/specific_learning/zero-shot/prepare.sh

Run evaluation for model with FID strategy trained on both GDELT and WIKI in the general learning stage:

bash scripts/specific_learning/zero-shot/run-FID-GL-GDELT-WIKI-130k-RID.sh

Low-resource setting

Data preparation:

bash scripts/specific_learning/low-resource/prepare.sh

Run training and evaluation with FID strategy on the earliest 5% training facts of ICEWS14:

bash scripts/specific_learning/low-resource/run-ICEWS14-05-FID.sh

Dataset Statistics

Dataset	Schema	# Entities	# Relations	# Train facts	# Valid Facts	# Test Facts	Time Granularity
ICEWS14	CAMEO	7,128	230	74,845	8,514	7,371	1 day
ICEWS18	CAMEO	23,033	256	373,018	45,995	49,545	1 day
YAGO	YAGO	10,623	10	161,540	19,523	20,026	1 year
GDELT	CAMEO	7,691	240	1,734,399	238,765	305,241	15 min
WIKI	Wikidata	12,554	24	539,286	67,538	63,110	1 year