rllm-team/rllm: relational tables as modules, packaged for docs only
Pytorch Library for Relational Table Learning with LLMs.
At a glance
- What is it?
- A PyTorch library that breaks GNNs, LLMs and TNNs into standardised modules so relational table models can be combined, aligned and co-trained. The packaging file states it exists for documentation, declares no dependencies, and the requirements file pins transformers to a ceiling that predates the LangChain release it requires.
- Who is it for?
- This fits a research group reproducing relational table learning papers from the example scripts rather than depending on a packaged install. Install from the requirements file rather than the package metadata, resolve the transformers ceiling before anything else, and expect the two unchecked todos, GPU or TPU optimisation and prompt optimisation, to be open work rather than shipped features.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on October 11, 2026, and from our analysis. They are not legal advice.
Editorial analysis
The only level-one heading is inside an HTML comment
The first line of the README wraps its title, documentation link and paper link in an HTML comment, so the page renders with no heading at all. What appears instead is a four-column link table carrying Documentation, Blog, Paper and Slide. The Blog points at rllm-project.github.io while the repository lives under the rllm-team organisation, and the Slide link is a PDF hosted on a personal site rather than in the repository. The paper is arXiv 2407.20157, and the documentation is hosted on readthedocs, matching the readthedocs.yml file at the top level. None of that makes the project unusable, but it does mean the README opens with a comment node before any content, which is the first thing a reader or a scraper meets.
The packaging file declares no dependencies and says why
The setup file carries a comment on its first line reading that it is only for docs so far. Everything below it is minimal. The distribution is named `rLLM`, the version is the string `0.1`, the package list is the single entry `['rllm']`, and `install_requires` is an empty list with a blank line inside the brackets. So a plain package install of rLLM pulls in nothing, and the twelve libraries the library actually needs live only in requirements.txt. The package list is also a hardcoded single name rather than a discovery call, which means any subpackage beneath `rllm/` would be left out of a built distribution. Combined with the comment, the file reads as scaffolding for the documentation build rather than a supported distribution path.
requirements.txt pins transformers below the LangChain it depends on
The dependency file is where the real constraints live, and one line dominates. `transformers<=4.30.0` is a hard upper bound, while the same file requires `langchain==1.0.2`, `langchain_core==1.0.1` and `langchain_community==0.4.1`. Those are three different generations of the LangChain stack, and the ceiling on transformers sits well below the releases the pinned LangChain line was built for. `torch` is bounded the same way with `torch>=2.1.0,<=2.3.0`, so there is no path to a current PyTorch either. Everything else is left open: `numpy`, `pandas`, `scipy`, `tqdm` and `pyarrow`, plus `networkx==3.0` and a `scikit_learn` entry written with a word separator where the published distribution name uses a hyphen.
The recommended starting example is the one method with no entry
The How to Try section points at BRIDGE, and the run instructions are three lines of comment plus one command:
# cd ./examples/bridge
# set parameters if necessary
python bridge.pyBRIDGE is the method a newcomer is told to start with, and it is also the only named method in the repository without an entry in the implemented methods list. That list highlights InRTL from KDD 2026, OGC from TNNLS 2024, ExcelFormer from KDD 2024, TAPE from ICLR 2024 and Trompt from ICML 2023, each with a paper link and an example link. No paper is cited for BRIDGE. The examples directory holds twenty-three entries and BRIDGE is one of only four that are directories, the others being llm_annotate, tape and transtab, while the remaining nineteen are single files.
Three organisations host the project's own follow-up work
Follow the links out of the repository and the owning organisation changes. This library is under rllm-team. The news feed points at rllm-project for the SJTUTables benchmark that accompanies the KDD 2026 paper, hosted at github.com/rllm-project/rllm_datasets. The January 2026 entry describes TLSQL, a SQL-like API for table learning, with its paper, its code at github.com/tlsql-project/tlsql and a separate demo video. The blog and course material also live under rllm-project.github.io, and the contribution guide is a Google Drive file rather than anything in the tree. None of this is hidden, but it does mean a benchmark and an API that the README presents as part of the story are installed and versioned somewhere other than the library they build on.
The version in the news feed outruns the version in packaging
The release history is told in the news feed rather than in tags, and the repository lists no GitHub releases at all. The July 2024 entry announces rLLM v0.1 along with the documentation. The January 2025 entry announces rLLM v0.1.1, framed as improving uniformity between Transform and Convolution operations. The packaging file, meanwhile, still declares `0.1`. So there are two places claiming a version, they disagree, and neither is anchored to a tag. The rest of the timeline is external validation rather than software movement: a KDD 2026 acceptance, a Snowflake approval in November 2024 with its own arXiv entry, a first prize for students in an open source security incentive programme in October 2025, and coverage in MIT Technology Review in December 2024.
The two unchecked todos are the ones that would change your throughput
The todo list has three completed boxes and two open ones. Code structure optimisation, support for more TNNs and large-scale RTL training are all marked done. GPU and TPU optimisation is unchecked, as is LLM prompt optimisation. Both open items sit directly on the hot path for anyone running this at scale, and the library is described as handling large fan-out graph data, so the gap between large-scale training being marked complete and GPU optimisation still being open is not explained on the page. The repository also carries a `test/` directory with no test command given anywhere in the README, and a `.flake8` file as the only lint configuration, with no companion tox, setup or packaging config file at the top level.
Foreign keys between tables are the unifying idea
The stated purpose has two halves. The first is breaking down state-of-the-art GNNs, LLMs and TNNs into standardised modules. The second is enabling new models through what is described as a combine, align and co-train sequence. The unifying claim is that different graphs, social networks, citation networks and e-commerce networks, can all be handled by treating them as multiple tables linked by foreign keys, which is what makes one interface cover many graph shapes. Around that sit the compatibility claims: integration with LangChain and Hugging Face transformers, three new relational table datasets shipped for model design, and over fifteen GNN and TNN models available standalone or as building blocks. The project is maintained by students and teachers from Shanghai Jiao Tong University and Tsinghua University, and feeds two named SJTU courses.
Editorial conclusion
This fits a research group reproducing relational table learning papers from the example scripts rather than depending on a packaged install. Install from the requirements file rather than the package metadata, resolve the transformers ceiling before anything else, and expect the two unchecked todos, GPU or TPU optimisation and prompt optimisation, to be open work rather than shipped features.
Frequently asked questions
How do I run an rLLM example?
The How to Try section starts from the BRIDGE example: change into `examples/bridge`, adjust parameters if needed, then run `python bridge.py`. Twenty-three example entries are present, with nineteen single files and four directories including bridge, tape, transtab and llm_annotate.
Does installing the rLLM package bring its dependencies?
No. The packaging file declares an empty `install_requires` list and carries a comment saying it is only for docs so far, and it hardcodes a single package entry. The actual dependencies, including LangChain, torch and transformers, are listed only in requirements.txt.
Why does rLLM cap transformers at 4.30.0?
The requirements file pins `transformers<=4.30.0` while also pinning `langchain==1.0.2`, `langchain_core==1.0.1` and `langchain_community==0.4.1`, so the ceiling sits below the LangChain generation it requires. Torch is bounded as `torch>=2.1.0,<=2.3.0`, with the remaining dependencies left open.
Which rLLM methods are listed as implemented?
Highlighted entries are InRTL from KDD 2026, OGC from TNNLS 2024, ExcelFormer from KDD 2024, TAPE from ICLR 2024 and Trompt from ICML 2023, and the library is described as including over fifteen GNN and TNN models. BRIDGE, which the quickstart tells you to run, is not among them.
Where are the rLLM datasets and the TLSQL API kept?
Neither is in this repository. The SJTUTables benchmark sits at github.com/rllm-project/rllm_datasets, and TLSQL, the SQL-like API for table learning announced in January 2026, sits at github.com/tlsql-project/tlsql, while this library is under rllm-team.
Is GPU optimisation finished in rLLM?
Not according to the todo list, where GPU and TPU optimisation and LLM prompt optimisation are both unchecked. Code structure optimisation, support for more TNNs and large-scale RTL training are marked done, and the repository lists no GitHub releases.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/rllm-team-rllm)