GraphGPT
Graph instruction tuning that lets an LLM read graph structure: text-graph grounding plus a lightweight alignment projector.
About
GraphGPT, from HKU's Data Intelligence Lab and Baidu (SIGIR 2024), is an early and influential attempt to give LLMs a native sense of graph structure. It uses a two-stage recipe: text-graph grounding that links textual descriptions to graph tokens, then instruction tuning through a lightweight alignment projector while the LLM and graph encoder stay frozen. The result is a model that can do node classification and link prediction, including zero-shot on unseen graphs, by reading graph structure as tokens.
The paper reports superior generalization over GNN baselines in both supervised and zero-shot settings, and the method has since become a standard reference in text-attributed graph research. Code and weights are on the HKUDS GitHub, with a project page and demo. The downsides are age and packaging: the LLM backbone is from the Llama-2 era, there is no Hugging Face distribution, no clear weight license was found, and the headline numbers are author-reported. Still, as the canonical graph-instruction-tuning baseline, it belongs in the catalog.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported gains on node classification and link prediction vs GNN baselines (paper, SIGIR 2024); no independent eval found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER metrics, metrics-v1). Evidence: arXiv 2310.13023 (Oct 2023), published SIGIR 2024, authors HKU + Baidu; github.com/HKUDS/GraphGPT (verified live) hosts code and weights; project page graphgpt.github.io; still cited as a baseline in 2025 text-attributed-graph work (TGRB). Strengths: canonical graph-instruction-tuning method, two-stage grounding+projector recipe, zero-shot graph generalization. Docked for: (1) all performance claims author-reported; (2) Llama-2-era backbone, no HF distribution; (3) no explicit weight license found. License: none declared.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · arXiv 2310.13023 (v1 2023-10-19), published SIGIR 2024; authors HKU Data Intelligence Lab + Baidu
- · github.com/HKUDS/GraphGPT verified live (curl 200, Oct 2026) with code and weights; project page graphgpt.github.io
- · Canonical graph-instruction-tuning baseline, still cited in 2025 text-attributed-graph work (TGRB)
- · Performance claims author-reported; Llama-2-era backbone; no HF distribution; no explicit weight license found