THU-KEG/SAEVerbalizer-27B

Partial fine-tune vs google/gemma-3-27b-it. Partial change: 68% of the weights changed (attention, mlp, norms), in 45 of 62 layers.

Tensors changed: Embeddings 0/4, Norms 51/484, MLP 135/294, Attention 180/464, other 0/1

2 derivatives checked

• mradermacher/SAEVerbalizer-27B-i1-GGUF: Quantization (fewer layers)
• mradermacher/SAEVerbalizer-27B-GGUF: Quantization (fewer layers)

On the Hub · Downloads on Model Pulse