Nobody asked for LLM changes. Nobody. First, Llama was doing fine. Then, you get us Deepseek. Which was meh, but it had its moments.

And when we were finally getting used to it, you change the model AGAIN.

What are you on to?

  • tlgklxzEnglish
    2·
    22 days ago

    Who told you the model is changed from Deepseek to X?

      • tlgklxzEnglish
        1·
        22 days ago

        That’s just a guess. The model is hallunicating heavily as always and this time it thinks it’s a Gemma model. Likely gemma 2 not even 4. Due to it says it’s local time is 21-22 may 2024 around 10:40 Am all the time.

        For Gemma 2 it’s right. But the knowledge cut off is around 2023. Not 2024. So it’s not even gemma but makes it more likely Mistral / Llama. But it does use lot’s of bolding, which is not Mistral nor Llama like.

        So there is only one option left… Deepseek as usual, the great identity thief. So no, no model change but there is a problem with system prompts leaking into context window. So most likely what we are having is the dev tweaking the model config and made a small mistake that caused bunch of bigger mistakes. Now we are suffering due to it. So calm your horses.

        Edit: WE don’t even know if it’s deepseek btw. Our own research results in Deepseek or heavily distilled model not even a base deepseek model. But we don’t know it. It uses deepseek tokenizer because it’s easy and free to use, which confuses some other models too.