AI safety · alignment · language models
The **Data Science & AI Lab** at [EPFL](https://www.epfl.ch), led by [Robert West](https://dlab.epfl.ch/people/west/), works on **AI safety and alignment**: how language models acquire values, how those values can be installed early enough to hold, and how to tell whether they actually did. Alongside that we study what models represent internally and how to make them more efficient. This Hub hosts the models, datasets, and benchmarks from our public releases. ## Research projects | Project | Paper | Code | On the Hub | | --- | :---: | :---: | :---: | | **SPP** โ installing an assistant persona from token zero | โ | [GitHub](https://github.com/epfl-dlab/model-raising-data) | [๐๏ธ dlab-spp](https://huggingface.co/dlab-spp) | | **zip2zip** โ inference-time adaptive tokenization via online compression | [arXiv](https://arxiv.org/abs/2506.01084) | [GitHub](https://github.com/epfl-dlab/zip2zip) | [Models](https://huggingface.co/collections/epfl-dlab/zip2zip-models-6852ec90f3dacc02aa6a0dca) | | **JSONSchemaBench** โ benchmarking constrained decoding on real-world JSON schemas | [arXiv](https://arxiv.org/abs/2501.10868) | [GitHub](https://github.com/guidance-ai/jsonschemabench) | [Dataset](https://huggingface.co/datasets/epfl-dlab/JSONSchemaBench) | | **Llaza** โ pretraining data mixtures | โ | โ | [Collection](https://huggingface.co/collections/epfl-dlab/llaza-ft-v01-69f7bb69776cea9530672480) | **Synthetic Persona Pretraining** has its own organization โ [**@dlab-spp**](https://huggingface.co/dlab-spp) โ with all models, the pretraining and post-training data, and the evaluation benchmarks. Browse [all GitHub repositories](https://github.com/epfl-dlab) or visit the [lab website](https://dlab.epfl.ch).