Seeking smart french-speaking model 10/30B to search large mythological texts

Models
1

I’m looking for an intelligent 10B to 30B model (GGUF for GPT4all) who speaks french for finding information in multiple, large texts, in the field of mythology.

I’ve got an i5 7600 with an old GPU (RX570 4GB) but I’ve got 48GB of Ram and I’m not in a hurry to get the answer, the most important thing is that the AI understands the mythological texts which are full of innuendo, allegories and unfamiliar words (characters, places, things, etc).

Can PrivateGPT, LM Studio or another be better than GPT4all (LocalDocs is efficient)?

Ministral-8B-Instruct-2410 is good for its size, but Mistral-Small-24B-Instruct-2501 or Qwen2.5-32B-Instruct aren’t much stronger. It takes 2 days to download so I can’t test many of them.

Online Claude and Deepseek are gifted, while others often have trouble understanding mythological texts, but I really do have a lot of PDF, so online isn’t a very interesting solution, unless it’s possible to search for interesting passages of text offline and send them to Claude or Deepseek.

I think I make fairly accurate prompts.
If you have any advice on GPT4all settings or anything else (I’ve already tried most of them).

1 Like
2

Models with excellent multilingual capabilities, such as Aya and Gemma, may be recommended because they have a high level of adaptability to unfamiliar languages.

In addition, I think that chatbot models for coding have the potential to understand unfamiliar terms to some extent as “that kind of thing” in order to understand the code. I will introduce some excellent ones.

Also, even if you can’t use it for the actual production of processing local PDFs, if you’re just trying out the model’s characteristics, you might be able to get a sense of how the model behaves before downloading it by using Spaces.

Based on the considerations and requirements, here is the recommendation for the most suitable AI models for analyzing mythological texts in French:

  1. Cohere’s Aya Expanse (11B parameters):

    • Strengths: Multilingual capabilities, including French; state-of-the-art performance across 23 languages, suitable for complex and nuanced texts like mythological content.
    • Compatibility: While primarily in GGUF format, its performance and size make it a strong candidate if GGUF compatibility is achievable.

  2. Mistral-Small-24B-Instruct-2501:

    • Strengths: Offers improved capabilities over smaller models, suitable for detailed analysis and inference in mythological contexts.

  3. Qwen2.5-32B-Instruct:

    • Strengths: Larger model size enhances handling of complex texts, ideal for extracting information from intricate mythological narratives.

  4. Alpaca-30B-Instruct:

    • Strengths: GGUF compatible, providing a balance between performance and efficiency, suitable for local setups.

  5. Falcon-30B-Instruct:

    • Strengths: High performance and GGUF compatibility, offering a robust solution for detailed text analysis.

Conclusion: Cohere’s Aya Expanse 11B is recommended as the top choice due to its multilingual capabilities and performance. For GGUF compatibility, models like Alpaca-30B-Instruct and Falcon-30B-Instruct are ideal. Consider local resource capabilities when selecting between these options.

3

Thanks for the answer, I’ll try that, I didn’t know “spaces”.

I’m also going to try mistral thinker, it’s more for role play but mythology is full of characters and story.

I forgot to ask, but I also have PDFs in English. What models are there for that?

Et pour les francophones qui passe par là il y a des modèles optimisés pour le français (chocolatine, lucie, etc)

1 Like
4

All of the models are usually taught English first, so they are all generally good at English. The models larger than 2B are generally multilingual in some way.

Also, there are some people on Hub who are trained in role-playing models, so it might be a good idea to try following the released models.

some fine-tuned RP models

about PDFs