H-Neuron modulation? WRT "H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs"

[DXXS]

Given the new paper on identifying causes of 'Hallucination's produced by LLM's [ https://github.com/thunlp/H-Neurons, as explained in https://www.youtube.com/watch?v=1ONwQzauqkc ], it appears to me that it might likely be worthwhile to allow for models to be loaded or converted while modulating the affect of neurons identified as such. Their provided source is reportedly currently ~dependent on vllm (?), but is said to result in identifying those 'indices identified as H-Neurons'.

I think it likely that the most straightforward way to do so efficiently would be to allow for an ~affector multiplier (default 1.0?) to be provided in mlx.lm.convert which would allow for generation of a modified model(?).