Feedback on the Russian language
I played a bit with your model and have to say that the gguf HA version worked terribly with the Russian language. There were lots of repetitions while the regular Q5_KM worked perfectly.
As for Russian language proficiency it is acceptable but on average slightly worse than SicariusSicariiStuff/Impish_QWEN_14B-1M. With some characters it still writes in English even when the greeting is translated. However its creativity in Russian is noticeably higher than Qwen's.
Also to me it feels somewhat similar to Magnum v4 though I can't quite explain why.
I haven't tested the context handling. I rarely go beyond 12–16K.
What HA quant have you used? (these are experimental, still testing if its worth it, and on what cases)
Regarding the Russian, the training data was small, but I am glad to see there was an improvement, as vanilla Nemo is not that multilingual.
Regarding Impish_QWEN_14B-1M, the multilingual ability is thanks to the Qwen base model, not due the training.
Thanks for the feedback 👍🏻
I used Q5_HA.gguf with three different characters, it got stuck in loops when generating responses in Russian, while in English everything worked perfectly. I switched to Q5_KM and the issue went away.
I mentioned Qwen because finetuning often degrades multilingual performance — for example, in smaller models like Mistral. Qwen, for me, is the benchmark among smaller models when it comes to Russian language support.
Thank you for the well-tuned model!
I really like how it describes actions and nonverbal cues using uncommon and expressive vocabulary.