[8th March 2026 Update]: Upload a new model (overriding the original one) without adapter on lm_head. Lora adapter on lm_head invokes a subtle issue on Peft that leads to device mismatch when using multi-gpus.

This model replicates LatentQA activation decoder for Llama-3.1-8B-Instruct. The original release of LatentQA decoder is Llama-3-8B-Instruct-based. This model has the same explantion results on the "promote_veganism" example documented in the repo of original work.

Activation decoder to interpret Llama-3.1-8B-Instruct's internal activations with natural language descriptions.

Please refer to https://github.com/aypan17/latentqa

The LatentQA fine-tuning is not extensive. The decoder is only good at answering questions analogous to fine-tuning dataset.