New models, Google Gemini, some adjustements

ngruver / llmtime

https://arxiv.org/abs/2310.07820

MIT License

673 stars 157 forks source link

New models, Google Gemini, some adjustements #25

Open zaizou opened 7 months ago

zaizou commented 7 months ago

In this update,

the Google Gemini Pro API model was integrated. It's important to note that Gemini isn't currently available in European countries. If you want to use it, you'll need to use a VPN.
Replaced Ploty for Plotly in the demo.py file to offer a more user-friendly graphical interface.
Added two draft models based on the Mistral API, specifically customized for finance.
made demo.py lightweight for faster generation. I chose to prioritize direct generation over working on hyperparameters.

Please note that the NLL calculation is not yet implemented for google, also for the past mistral model.

zaizou commented 7 months ago

Are you we okey ?

ngruver commented 7 months ago

Hello,

I think I will need to update the readme (beyond what is proposed in the PR) to make it more clear now that there are many model options. I can take a look this weekend.

zaizou commented 7 months ago

Okey, Maybe also working on the architecture (OOP), so when a company modify or add a model (API) (Mistral large, and gemini ultra for instance) we don't have to code. It will be interesting that we discuss about the feature evolutions if it will some. Maybe some integration with other research and thesis. Let me know if you're open for a collaboration I have some ideas 👍

ngruver commented 7 months ago

I'm reading through your PR and I noticed there might be an issue in the last set of changes. In the code for the mistral API, it looks like you are using the OpenAI tokenizers: https://github.com/ngruver/llmtime/blob/main/models/mistral_api.py#L33. Did you confirm that this makes? It probably makes more sense to use the huggingface mistral tokenizers, unless it is known that their API models use a different tokenizer.

ngruver commented 7 months ago

Other issues/comments:

What is the difference between "mistral-api" models and "mistral-api-stocks" models? Is it just a different prompt applied to the same models? There might be a simpler way to implement this, by adding a "sys_message" argument for example.
.vscode files should be removed from the PR
finbert_utils.py should also be removed, as I'm not seeing how it is used

zaizou commented 7 months ago

Hello, Of course all of your remark are good I wanted to pass the system message as a parameter but I didn't have much time, I thought also about making things OOP To two other points will be done soon

zaizou commented 7 months ago

Also for Mistral tokenizer ok Tell me if you wanna do it yourself