g-luo / generative_latent_prior

What is generative_latent_prior?

This PyTorch repo delivers the official implementation for training generative diffusion models on LLM activations, modeling their natural manifold. Load pre-trained weights from HuggingFace for Llama 1B/8B layers, generate realistic activations, or project edits back on-manifold via simple API calls like `load_glp("generative-latent-prior/glp-llama8b-d6")`. It tackles off-manifold activation steering that breaks LLM behavior, enabling cleaner interventions on activations from official PyTorch tutorials or datasets.

Why is it gaining traction?

Pre-trained models fit on a single RTX 4090 with under 24GB VRAM, and toy training runs in minutes on 1M activations—far easier than full diffusion setups. Built-in scripts handle scalar probing across 113 datasets and on-manifold post-processing for persona vectors, with seamless HuggingFace integration like official PyTorch Docker images. Developers grab it for quick experiments without activation caching hassles.

Who should use this?

Interpretability researchers probing LLM layers for features, safety engineers refining activation steering beyond naive additions, or anyone analyzing activations from Llama models via official PyTorch get started page flows. Ideal for steering experiments with persona vectors or evaluating generated activations against real ones.

Verdict

Promising for activation work, with strong README quickstarts, demo notebooks, and HF weights, but early maturity shows in 25 stars and 1.0% credibility score—test on small datasets first. Worth forking if you're in LLM internals.

(198 words)