kitft/nla-gemma3-27b-L41-av
kitft/nla-gemma3-27b-L41-av is a 27 billion parameter Natural Language Autoencoder (NLA) activation verbalizer (AV) model, fine-tuned from google/gemma-3-27b-it. It maps hidden-state vectors to natural language descriptions, serving as an interpretability tool for LLM activations. This model is designed to be used in conjunction with its paired activation reconstructor (AR) for analyzing and explaining internal model states, rather than as a general-purpose language model.
Loading preview...
Model Overview
This model, nla-gemma3-27b-L41-av, is the Activation Verbalizer (AV) component of a Natural Language Autoencoder (NLA) pair, derived from google/gemma-3-27b-it. It is specifically fine-tuned to translate a large language model's internal hidden-state vectors into natural language descriptions. Its counterpart, kitft/nla-gemma3-27b-L41-ar, reconstructs these descriptions back into vectors, enabling a unique approach to LLM interpretability.
Key Capabilities
- Activation Decoding: Maps residual-stream activations from block 41 of a Gemma-3 model to human-readable text.
- Interpretability Tool: Designed to explain what specific internal activations "mean" within an LLM.
- Paired Functionality: Intended for use with its reconstructor (AR) model to measure the fidelity of the natural language explanation.
Good For
- LLM Interpretability Research: Analyzing and understanding the internal workings of large language models.
- Activation Analysis: Gaining insights into the semantic content encoded in specific activation patterns.
- Debugging and Development: Potentially aiding in the identification of problematic or interesting internal states during model development.
Note: This model is not suitable for general-purpose language generation or instruction-following tasks, as its fine-tuning has repurposed it entirely for activation decoding.