You are browsing as a guest. Sign up (or log in) to start making projects!

(Re)Training an AI model on Tenstorent Hardware

By @Bjarnos

  • 1 Devlogs
  • 1 Total hours

I'm very interested in developing Neural Networks and Large Language Models (quite a different scale, I know 😅). I found out I have enough compute to retrain a Llama 3.1 8B model with Lora, so that's what I'm going to do! Will I succeed in making a study assistant for Dutch exams?
Open comments for this post

@Bjarnos on (Re)Training an AI model on Tenstorent Hardware

1h 24m 17s logged

Devlog 1: First Steps

Project Vision & Architecture

I want to build an educational AI assistant for Dutch high school students preparing for their final VWO (and HAVO) exams. I started this project because I think many students struggle to get detailed feedback while studying independently, and AI could help fill that gap. I also noticed a similar project already existing (https://aivoorleerlingen.nl/), but they put the study knowledge in context, rather than an actual database and retrained AI models. I hope that I got you a little excited :D.

To achieve this project, I’m going to use the AI architecture shown in the screenshot below.

  1. Training the AI: The model is trained on exam questions and answers so it learns how VWO and HAVO exam solutions are typically written.
  2. Providing study material: The AI can look up relevant curriculum information before answering, which helps it give more accurate explanations.

What is working already

I successfully built the first data preparation script. It can now process scraped exam questions and convert them into a format that can later be used for AI training. I don’t have any screenshots yet, but I’m planning to release the source as soon as the training loop is complete.

Next Actions & Milestones

  1. Acquire Textbook Data: Create and test prepare_rag_db.py to see if it can initialize the vector database.
  2. Complete the Tenstorrent Training Loop: Finish the training script by implementing extra steps.
  3. Establish Inference Pipeline: Build an evaluation script that connects the fine-tuned model with Qdrant vector retrieval, generating and comparing exam responses against the test set.

I would appreciate it if you want to follow my progress! If you want more explanations on a specific step, or if the next devlog should be in simpler language, let me know.

@Bjarnos

Devlog 1: First Steps

Project Vision & Architecture

I want to build an educational AI assistant for Dutch high school students preparing for their final VWO (and HAVO) exams. I started this project because I think many students struggle to get detailed feedback while studying independently, and AI could help fill that gap. I also noticed a similar project already existing (https://aivoorleerlingen.nl/), but they put the study knowledge in context, rather than an actual database and retrained AI models. I hope that I got you a little excited :D.

To achieve this project, I’m going to use the AI architecture shown in the screenshot below.

  1. Training the AI: The model is trained on exam questions and answers so it learns how VWO and HAVO exam solutions are typically written.
  2. Providing study material: The AI can look up relevant curriculum information before answering, which helps it give more accurate explanations.

What is working already

I successfully built the first data preparation script. It can now process scraped exam questions and convert them into a format that can later be used for AI training. I don’t have any screenshots yet, but I’m planning to release the source as soon as the training loop is complete.

Next Actions & Milestones

  1. Acquire Textbook Data: Create and test prepare_rag_db.py to see if it can initialize the vector database.
  2. Complete the Tenstorrent Training Loop: Finish the training script by implementing extra steps.
  3. Establish Inference Pipeline: Build an evaluation script that connects the fine-tuned model with Qdrant vector retrieval, generating and comparing exam responses against the test set.

I would appreciate it if you want to follow my progress! If you want more explanations on a specific step, or if the next devlog should be in simpler language, let me know.

Replying to @Bjarnos
0

Followers

Loading…