LiteLink

LiteLink is a lightweight research codebase for low-resource KG link prediction experiments, based on ULTRA-style inference and LLM-assisted relation prompts.

Project Structure

• config/: inference and pretrain configs
• script/: training / inference / few-shot / reporting scripts
• ultra/: model and task implementation
• llmoutput/: relation-text outputs used by prompting
• ingram_fewshot_ds/: few-shot dataset json files
• ckpts/: checkpoints used for inference
• results/: summary csv / markdown outputs

Environment

Recommended:

• Python 3.10+
• PyTorch 2.1+
• PyTorch Geometric 2.4+

Install dependencies:

pip install -r requirements.txt

Quick Start

1. Prepare datasets (do not commit large raw/preprocessed data to git).
2. Ensure checkpoint files exist under ckpts/.
3. Run a few-shot inference example:

python script/run_fewshot.py -c config/inductive/inference.yaml --gpus [0] --ckpt ckpts/ultra_3g.pth -d MedIngram:100

4. Build result summary:

python script/build_final_report.py

Notes for GitHub Upload

This repository is prepared as a code-focused version. Large generated artifacts, cached wheels, and local training outputs are removed before upload.

Suggested: keep large dataset files and intermediate training outputs outside this repository, or manage them with external storage.

Acknowledgement

Core model implementation references ULTRA and InGram related resources.