generative-latent-prior/frechet-distance-fineweb-50k

Llama Frechet Distance Reference Activations

This repository contains activation data accompanying the paper Learning a Generative Meta-Model of LLM Activations.

• Project page: https://generative-latent-prior.github.io
• Code: https://github.com/g-luo/generative_latent_prior

Quick Start

With this data, you can compute the representation Frechet Distance of generated activations from GLP. This repository contains real reference activations derived from FineWeb. Each activation is taken from a different document, at a random token position. The file llama1b.pt corresponds to Llama-3.2-1B Layer 07 activations. The file llama8b.pt corresponds to Llama-3.1-8B Layer 15 activations. GLPs are activation diffusion models useful for applications like on-manifold steering and sparse probing.

python3 glp/script_eval.py

Citation

@article{luo2026glp,
  title={Learning a Generative Meta-Model of LLM Activations},
  author={Grace Luo and Jiahai Feng and Trevor Darrell and Alec Radford and Jacob Steinhardt},
  journal={arXiv preprint arXiv:2602.06964},
  year={2026}
}