Add Chinese training materials

#1
by btaskel - opened

HellošŸ˜Š, I really like the model you created!! It's simply the best galgame model I've ever encountered! I don't know much about AI, but are you considering adding some Chinese training materials? šŸ¤”It would be great if this could improve its Chinese language proficiency.ā¤~
ciallo~

Owner

Hi there,

Thank you so much for your kind words and for your interest in the model! šŸ˜Š I'm really glad to hear youā€™re enjoying it.

At the moment, Iā€™m focusing on Japanese and English data, as those are the languages Iā€™m more familiar with, which helps me better refine and adjust the model.

Unfortunately, since Iā€™m not as familiar with Chinese, I wouldnā€™t be able to accurately assess the quality of the data or how useful it might be for the model. For now, I donā€™t have plans to add Chinese data, but I truly appreciate the suggestion!

spow12 changed discussion status to closed

Hello, I really like your Chatwaifu LLM modelā¤, but as a non-professional in AI, I have a question that I find a bit confusing: Qwen 2.5 14B and Mistral Nemo 12B both use the Apache 2.0 license. In the "Open LLM Leaderboard Evaluation Results," Qwen 2.5 has a higher score. Does this mean it would be better to use Qwen 2.5 as the base model in Chatwaifu? šŸ˜Š

Qwen 2.5: MMLU-PRO (5-shot) 48.85
Mistral Nemo: MMLU-PRO (5-shot) 28.91

Owner

In general, the Leaderboard Evaluation Results assess the overall performance of large language models (LLMs) in English.
So..., you may be rightā€”based on these results, Qwen 2.5 appears to demonstrate stronger general performance compared to Nemo.

However, I have two reasons for preferring Mistral Nemo:

  1. Chatwaifu is intended to act as an AI waifu, focusing on specialized rather than general tasks.
    In my experience, Mistral Nemo performs better for Japanese roleplaying, which is the primary focus of my research, compared to Qwen 2.5.

  2. The evaluation results primarily measure only English performance not chinese and japanese either.
    So, i use these results as a reference to confirm that the LLM maintains satisfactory general chat capabilities after fine-tuning to act as an AI waifu.

In summary, I prefer Mistral Nemo due to 'my personal experience' with it in these specific applications.

In general, the Leaderboard Evaluation Results assess the overall performance of large language models (LLMs) in English.
So..., you may be rightā€”based on these results, Qwen 2.5 appears to demonstrate stronger general performance compared to Nemo.

However, I have two reasons for preferring Mistral Nemo:

  1. Chatwaifu is intended to act as an AI waifu, focusing on specialized rather than general tasks.
    In my experience, Mistral Nemo performs better for Japanese roleplaying, which is the primary focus of my research, compared to Qwen 2.5.

  2. The evaluation results primarily measure only English performance not chinese and japanese either.
    So, i use these results as a reference to confirm that the LLM maintains satisfactory general chat capabilities after fine-tuning to act as an AI waifu.

In summary, I prefer Mistral Nemo due to 'my personal experience' with it in these specific applications.

Oh! I understand what you mean now. Thank you for taking the time to answer my question! Wishing you a wonderful day.šŸ˜‰

Sign up or log in to comment