Frame2KG aggregates the data, models, and tooling introduced in Frame2KG-YC2: A Synthetic Dataset, LoRA Adapters, and an Evaluation Toolkit for Frame → Graph. The goal is to provide a single launch point for reproducible frame-to-knowledge-graph research focused on interpretable, on-device robotics.
- Dataset: Frame2KG-YC2 synthetic corpus (HuggingFace Repo)
- Synthetic Dataset Creation Scripts (Dataset Generation)
- Training pipeline: Frame2KG-Trainer for Qwen2.5-VL LoRA adapters (Frame2KG Trainer)
- Evaluation toolkit: Deterministic frame2kg-eval matcher and metrics (Frame2KG_Eval_Benchmark)
- Adapters: Released LoRA checkpoints (3B/7B; QKVO ± GateProj) HuggingFace Collection)
- Ablations: Effect of adding MLP gating (GateProj/Up/Down) on Node/Edge F1µ and IoU (read the ablation summary)
- Paper: Frame2KG baseline reports and comparison results (coming soon)
- Appendix details (Appendix.md)
2026-04-13:
- Improved evaluation efficiency (shared embedding model, reduced per-frame overhead)
- Standardised evaluation defaults for node text fields (
label, attributes)
Note on evaluation:
The evaluation toolkit useslabel, attributesas the default node text fields.
For exact reproduction of the LREC 2026 paper results, please refer to the evaluation repository documentation.
@inproceedings{watson2026frame2kg,
title = {Frame2KG: A Benchmark and Evaluation Toolkit for Interpretable Frame-to-Graph Generation},
author = {Watson, Lewis N. and Strathearn, Carl and Mitchell, Kenny and Yu, Yanchao},
booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
month = {May},
year = {2026},
pages = {10912--10926},
address = {Palma, Mallorca, Spain},
publisher = {European Language Resources Association (ELRA)},
editor = {Piperidis, Stelios and Bel, Núria and van den Heuvel, Henk and Ide, Nancy and Krek, Simon and Toral, Antonio},
doi = {10.63317/4ys6kofrzoc5},
url = {https://doi.org/10.63317/4ys6kofrzoc5}
}For questions or collaboration inquiries, reach out to l.watson{at}napier.ac.uk.
Replace {at} with @
