Neural DareDevil-8B 😈: The fastest LLama3 8B Finetune + Merge on earth!

Calling all Large Language Model enthusiasts! This video showcases Daredevil-8B, the new champion of 8B parameter models on the Open LLM Leaderboard. Through a technique called abliteration and DPO fine-tuning, we've achieved an uncensored model that surpasses even the powerful Llama 3 Instruct 8B across all tested benchmarks! See how Daredevil-8B pushes the boundaries of large language models and stay tuned for further exploration of its capabilities!

Tell us what you think in the comments below!

-----------------

This video contains affiliate links, meaning if you click and make a purchase, I may earn a commission at no extra cost to you. Thank you for supporting my channel!

My 4090 machine:

Tech I use to produce my videos:

[GerryPrompt]

No way! Had no idea you could merge llama 3 like this! Can't wait to see more from this dev 😮

[maximelabonne]

Hey, thanks for reviewing my model, I enjoyed the video! Also happy you found it interesting.

I'm not familiar with the inference endpoint you used to run the model, but it might be an issue related to the chat template (or other parameters). The model does correctly shut up when you prompt it in LM Studio with Llama 3 or Llama 3 V2 presets. :)

[Raskoll]

A Deep dive on how merging works would be cool. Like the maths/code behind it. Would be hard tho

[-dy]

so the merge process was kinda like yam peleg's "experiment" series? competing models for scores and from experimentation data, merge the models which tend to give a performance increase and remove the models which tend to draw down the performance.

[supercurioTube]

Oh man I'm excited watching this.
I'm experimenting with abliterated models for my project to avoid any false refusals, then evaluating the output for harmful advice in a pipeline. I'm hoping it will be a more granular approach to safety than something built-in the model itself.
So I'm interested in this model if it's finally one better than Meta's.
As a merge, I'm assuming it might be contaminated tho and then the benchmark results could be meaningless... okay watching now

[eck]

I fear that those scores are directly integrated into the training process, which is essentially minimize loss to get the better score, by basically cheating and not generalizing at all.

[TheSuperStroker]

Before you close the window or refresh, the first 2 1/2 minutes is one still image. Nothing is wrong with your hypnotic rectangle. It's just… You know. one still image. Sorry about your ADD

[jasonreviews]

i just tried it. it's pretty broken it kept spitting out python code

[aneeshprasobhan]

Am I the only one that wants to know the background video tool that mapped 3D space using drones ?

[jeff_]

This model doesn't seem to be uncensored.

[Markus-rg]

what benefits do mmlu specific models give?