Name: ToRA
Uploaded: 2025-02-18T13:23:15.770Z
Duration: 2 min 33 s
Description: A tool-integrated AI agent designed to solve complex mathematical problems by interacting with external tools.

tora

A tool-integrated AI agent designed to solve complex mathematical problems by interacting with external tools.

ToRA

In this video I'm going to show you how you can locally install Tora which stands for a tool integrated reasoning agent. 

Now what is Tora? Torah is a series of tool integrated reasoning agents designed to solve challenging mathematical reasoning problems by interacting with tools for example computation libraries or symbolic solvers. 

This Tora series models seamlessly integrate natural language re with the utilization of external tools thereby mixing the analytical process of language and the computational efficiency of external tool. 

Right now Tor code 34 billion model is the first and only open-source model to achieve over 50%.

Accuracy on the math data set which significantly outperforms even GPT Force Chain of Thought result and is competitive with GPT Force solving problems with programs. They have also open-sourced their code and models and then they are hoping to achieve more breakthrough in this. 

I will drop the link to this GitHub repo where you can read more about all this benchmarking, their source code and a lot of other things including their research paper, their website, Hugging Face model and a lot of other things. 

Also, these models, the Tora models I mean, are trained on Torah Corpus 16k which contains tool integrated reasoning.

Trajectories of math and GSM 8K from gbt 4. They also use imitation learning or sft to fine-tune the models and then apply their proposed outspace shipping to improve tool integrated reasoning behaviors. You can read more information about it in the paper. 

I'll be using the blogs Tora 7B quantized version which is present on this hugging phas page, and I will also share the link in video's description. The file size is quite cool, around 4 gig, so it should be good. You can install this model either on Windows, Linux, Google collab, or even aw sagemaker notebook. 

I'll be using my trusted Google collab notebook for this, and I'm using the

T4 runtime which is a GPU one so make sure that you select the runtime type as t or GPU and similar ones on your awcc instance if you're using that or any other system. If you want to install it on Windows, make sure that you used Windows Subsystem for Linux in order to run the commands which I'm going to use. 

So let's get started. The first step we need to do is to install the prerequisite libraries which are Transformers and Optimum. So let's run them. Shouldn't take too long. Let's wait for it to finish. Both are done. In the next step, let's get clone the gptq library because this is what is needed to use the quti version we have.

Cloned it in the next step. Now let's CD to that directory which is our Auto gbt q1. We are in that directory. Let's check out the stable version from this GitHub repo which is version 0.4.2. That is nice and now let's install it which is simply our familiar command pip install dot. This takes around 2 minutes so let's wait for it to finish. 

Auto gptq is installed now. In the next step let's download the model from the blog score conted version from hugging phase which is Tora 7B V1. Then use the auto tokenizer to generate the token. So let's run it and the model size as I mentioned earlier is around just over 3 gig I guess.

So shouldn't take too long. Let's wait for it to start. The first it is download downloading the config file. And then as you can see around 4 gig 3.9 gig model is being downloaded. Let's wait for it to finish downloading. The model is downloaded now. 

Let's give it a complex math problem as a prompt. In this code chunk I asking it bit of a complex math equation that if 3x - Y is = 12 what is a value of 8 into x 2 into Y. And then I'm giving it a prompt template which is required by this model. And then I'm using autot tokenizer to generate inputs and then feeding it to the model which generates output. And then we feed it into the.

Pipeline to print out the response of the model so let's run it and see what happens. I will let it run in real time so that you can also um see the response time. So you can see how quick it was and it has given us uh not only the solution of it but also uh program around it. It has run it and it has given us this output and there you go. 

Now um this model as I said earlier is primarily gate towards solving the complex math problems. So feel free to give it complex problems like um exponential and even your set is set problems uh set exam uh problem so that and see how it goes. But if you want to install it now you have the steps.

And how to do it you can install it in Linux in any version of Linux even if it is hosted in WSL. I hope that you enjoyed it. Um if you already have tried it out let me know how you go. And if you have any questions please ask. And if uh you like the content please consider sharing AI Agent Store with colleagues. Thank you.

Note, that there may be slight differences in text and video. Watch video for the original content.

ToRA Video