Aura/Luna

#1
by EloyOn - opened

I noticed that you always merge Aura and Luna L3 LoRAS. It changes too much the personality if you only use only one of them? Jeiku did too, with Orthocopter L3.

Original Aura L3, simply had Aura and Bluemoon.

I've been keeping an eye on Daredevil too, great choice for a merge!

Owner

All of the merges with "SOVL" are attempts to clone the style of ResplendentAI/SOVL_Llama3_8B but including the latest models, bigger merges and including OAS/Abliteration. I don't really have a methodology it's mostly trial and error.

I just spotted daredevil this evening and messed around in LM-Studio. It was the first model with abliteration that was at least equally as smart as the original llama-3-instruct while still being willing to answer messed up questions in instruct mode!

Yeah, abliterated models are a little dumber that non-abliterated ones. It seems NeuralDaredevil-8B-abliterated is slightly better than normal Daredevil-8B though, at least looking strictly at the benchmark.

https://huggingface.co/mlabonne/NeuralDaredevil-8B-abliterated

Owner

Yeah, abliterated models are a little dumber that non-abliterated ones. It seems NeuralDaredevil-8B-abliterated is slightly better than normal Daredevil-8B though, at least looking strictly at the benchmark.

https://huggingface.co/mlabonne/NeuralDaredevil-8B-abliterated

Missed the neural version, a merge is running now 😸

Owner

I imagine the neural version will be a little less unhinged, probably better to chat to for most scenarios. It would hopefully put a little bandaid over the damaged caused by abliteration

I imagine the neural version will be a little less unhinged, probably better to chat to for most scenarios. It would hopefully put a little bandaid over the damaged caused by abliteration

Uncensoring usually lowers the smarts of the model, it seems. Less unhinged but smarter is probably a safe bet for this NeuralDaredevil version.

Owner

Sign up or log in to comment