AtlasKV
About this tool
The official implementation of an ICLR 2026 paper encoding billion-scale knowledge-graph triples as key-value vectors for model knowledge augmentation within 20GB of GPU memory. Includes KG2KV data preparation, training, memory and accuracy evaluation scripts, and checkpoints trained on ATLAS-Wiki and ATLAS-CC.
- Feature 1: KG2KV converts graph triples into QKV training data.
- Feature 2: Support for billion-scale ATLAS-Wiki and ATLAS-CC graphs and synthetic data.
- Feature 3: Choose text-embedding-ada-002, text-embedding-3-large or all-MiniLM-L6-v2 encoders.
- Feature 4: Scripts evaluate memory use, knowledge accuracy and generation relevance.
- Feature 5: Checkpoints trained for 3K steps are available.
Resource information
- Maintainer
- HKUST-KnowComp, Associate Professor Yangqiu Song's group.
- Type
- Tool / library
- Category
- Knowledge graphs & graph computing
- Runs on
- Tool or library
- Who it's for
- Researchers studying model knowledge augmentation, graph–model integration and lower-memory methods, including single-GPU reproduction and comparisons.
- License at time of listing
- MIT
- Source
- GitHub
