Hello there,

I’m not sure if this is the right sub to ask this question, so please tell me if there’s a better sub to ask this question.

I’m a student and me and my team have the assignment to make a chatbot for our university. We need to make a chatbot that can help other students find information about their course. We will get our data from manuals of multiple universtity websites (as pdf). This data will be turned into Q&A data using ChatGPT 4.

However, we are struggling to find a pre-trained LLM that fits our assignment. We’ve researched T5, BERT and GPT-2 but our teacher was surprised those were the models we researched, since there are more popular and newer models. Our chatbot must be in Dutch, but we can translate so the LLM doesn’t need to be trained on Dutch data. The LLM can’t be too big, because we don’t have the hardware for very large models.

We are currently looking at openchat and falcon, both with 7B parameters. Are these good options or does anyone have any tips for better LLMs?

  • rainbow3@alien.topBEnglish
    1·
    10 months ago

    This is much easier than you think. Instead of retraining look at Retrieval Augmented Generation. This creates a database of your documents that can be queried for relevant passages. Then any requests plus relevant sources from your documents are sent to the LLM to formulate a response. You can use your own data; it provides source references; and can add new documents as required with zero retraining.

    Using llamaindex or Langchain this requires < 50 lines of code. One line change to use a different LLM provider. Alternatively openai have launched GPTS which does it completely code free.

  • lifesthateasy@alien.topBEnglish
    1·
    10 months ago

    I think you’re going to have a hard time completing this assignment. Just based on the fact you were unable to read the sub’s rules. I don’t think whatever we link you you’ll be able to read the readme of.

  • Main_Path_4051@alien.topBEnglish
    1·
    10 months ago

    This is a kind of RAG augmented project or llama project with embeddings . Not really complicated. have a look at langchain too, altough this may not be very efficient, that is a first approach.