watsonx.ai

 View Only
  • 1.  query regarding adapting watsonx foundation model to custom data.

    Posted Thu April 25, 2024 10:15 AM

    Hey, I am the tech lead from our company. We are using IBM WatsonX as a core component. I have looked around the docs and many articles, but I do not yet have a clear idea, I just need pointing in the right directions. 

    How do we take up a foundational model, dump in pages upon pages from our database to train it and then save it. Afterwards in the future, when the same model is prompted, it should answer using its its llm capabilities, while using the data and information it was trained upon by us.

    I know that finetuning and embeddings exist for similar use cases. In case of finetuning, please do correct me if I am wrong, but I do not think dumping in information would be possible, we'll have to have good prompt answer sets right? In case of embeddings, even though embeddings can selectively pick up relevant data, and put only those within the model context, In our use case even that might not be enough.

    We are currently using the embeddings approach, but to have a wider range of context, we want the model to be trained upon our whole database, but we dont want to train it from scratch.

    tldr: how do we dump in tons of data (exceeding context length by a very huge huge margin) on a foundational model, and have it use that data when it answers in the future? we don't want to use embeddings, if its possible using finetuning, please guide how.

    Thank you.


    #watsonx.ai

    ------------------------------
    swayam shree
    ------------------------------


  • 2.  RE: query regarding adapting watsonx foundation model to custom data.

    Posted Fri April 26, 2024 02:32 PM

    Hi Swayam,

    Sounds interesting. 

    From my knowledge, Fine tuning is more suited if you want to tune the model to perform a specific task that is needed for your use custom case.
    Embedding is more suited when you have a knowledge base, that can be referred to , so the hallucinations will come down. We can connect offline, if you are interested. I can look in to your use case and give my thoughts for your specific case. - Arvin, +91 96000 38297





    ------------------------------
    Arvin Subramanian
    ------------------------------



  • 3.  RE: query regarding adapting watsonx foundation model to custom data.

    Posted Mon April 29, 2024 05:41 AM

    Swayam, as watson.ai is ne to most of us I see many questions going the same direction. If you got useful input from Arvin I would be interested as well. Personal mailbox here is one way. There we can exchange contact data. MS meeting is one way to work on the lab.

    My thoughts are:

    • personal documents need to be classified to be used for effective training
    • the model parameters and the model itself should be modified to prefer your own documents over generative ai talk
    • embedded approach is too specific cause it uses your own parameters derived from you environment (same approach as using e.g. salesforce integration service or similar)


    ------------------------------
    [Karl] [Jaeger] [#ibmchampion]
    [QRadar Specialist]
    [cnag]
    [Siegen] [Germany]
    ------------------------------



  • 4.  RE: query regarding adapting watsonx foundation model to custom data.

    Posted Tue April 30, 2024 05:17 AM

    Depending on your use case, you might use Retrieval Augmented Generation i.e. index your data into a vector index in watsonx Discovery SW or IBM Cloud Databases for Elastic Search Platinum or Elastic Search Enterprise SW, and use the Q&A with RAG Accelerator for watsonx.ai with that. For watsonx.ai aaS on IBM Cloud you can find the Q&A with RAG accelerator on the watsonx Resource Hub. For watsonx.ai SW you could contact paul.kilroy@ie.ibm.com to get the Q&A with RAG accelerator version for that.



    ------------------------------
    Thomas Schaeck
    ------------------------------



  • 5.  RE: query regarding adapting watsonx foundation model to custom data.

    Posted Wed May 08, 2024 09:25 AM

    Hi Thomas,

    Can you share the link to documentation on how to ingest data into vector db (watsonx Discovery SW or IBM Cloud Databases for Elastic Search Platinum or Elastic Search Enterprise SW).

    Thank you



    ------------------------------
    Amir Zahoor
    CTS Manager
    IBM
    (513) 817-7194
    ------------------------------