Chengheng's picture
Upload README.md with huggingface_hub
a0d501e verified
metadata
license: mit
library_name: pytorch
tags:
  - coconut
  - latent-chain-of-thought
  - sparse-autoencoder
  - qwen3

SAE on Coconut latent thoughts (Qwen3-1.7B)

SAE bottleneck trained on the frozen Stage-3 Coconut latent thoughts of Qwen3-1.7B, from the VQ-CoT: Discretising Latent Chain-of-Thought project (team RateLimit Achieved, EPFL CS-552). The language model is frozen; only this bottleneck is trained on the dumped latents (385,620 x 6 latent thought vectors, GSM8K).

  • Checkpoint: sae.pt
  • TopK SAE, dictionary 8x hidden, k=32 active features
  • Result: inserting this bottleneck into the frozen latent loop costs ~0 GSM8K test accuracy.