Orenguteng

/

Llama-3.1-8B-Lexi-Uncensored-GGUF

like 19

It's more compliant, got better reasoning, as for other factors of Llama 3.1 vs Llama 3 I have not been able to fully personally evaluate the model yet due to lack of time.

My first attempt got blocked. 3.0 is definitely less censored.

@askyforever Thanks for your input. This model has retained much if not all of it's intelligence and reasoning, therefore it might refuse some prompts, this can be circumvented by simply editing the first words of the response such as "Sure, to do XYZ" and continue generation. This method should work for most if not everything, however it would still trigger a refusal in the original non-censored model.

I will release an even more compliant version update soon, while trying to retain the original intelligence as the objective is.

If you can provide more info on the prompts tested, you can DM me on discord @disgrace6161

Edit:
Try follow up with "Thanks for the tips. Now answer the question." or something similar, it should resolve most issues for any prompts, if you don't prefer the described method of pre-filling. However, next release will be even more compliant.

Llama 3.1 is very bad in reasoning, logic and math. It's a worldwide problem that people are regretting... One step forward, two steps back... And we don't know the solution. Some blamed the pre tokenizer, but we really don't know why version 3.0 is better than version 3.1

🙏👍💥

@Dihelson I haven't fully evaluated Llama 3.1 due to lack of time. Since you experice that it's worse at reasoning, logic and math, can you show some example prompts where 3.0 answers correctly and 3.1 fails for reasoning and logic? Would be good to see for myself.

@Dihelson I haven't fully evaluated Llama 3.1 due to lack of time. Since you experice that it's worse at reasoning, logic and math, can you show some example prompts where 3.0 answers correctly and 3.1 fails for reasoning and logic? Would be good to see for myself.

Hi, Orenguteng, I'll later try to provide you with some tests I did on both, or try to do them again, because they are scattered throughout several conversations. But I can run them again and create a table of comparisons with questions and answers, but it will take some time, and today I'm very busy. Perhaps at the end of the day. Thanks. 🙏👍💥

@Dihelson Great thanks!