We couldn't load all Actvitity tabs. Refresh the page to try again.
If the problem persists, contact your Jira admin.
Uploaded image for project: 'translate5'
  1. translate5
  2. TRANSLATE-4438

Integrate Llama 3 as language resource analogous to GPT

Details

    • Improvement
    • Resolution: Unresolved
    • None
    • None
    • openai

    Description

      Requirement

      Llama should be integrated in translate5 as trainable language resource completely analogous to what we are doing at the moment with OpenAI and Azure GPT.

      • persistent training of models should be possible with prompts and exmaples
      • The new training lib features as developed in https://jira.translate5.net/browse/TRANSLATE-4369 should be possible to be used
      • Context matters. We should send as many segments as possible at the same time to Llama. We should count the tokens before sending the request for translation to Llama, to make sure, we send as many segments as possible for translation.
      • We should send the terminology to Llama, in the way, we do for GPT and be prepared, that we will implement also https://jira.translate5.net/browse/TRANSLATE-4401 and its sub-issues soon for GPT and for Llama
      • We should evaluate, if there is  a better RAG-based approach for enriching GPT with the terminology (and soon matches from the TM) than what we are currently doing
      • We should have in mind, that for quality estimation we will probably also need and use the training interfaces

      Llama model to be used as long as Llama 3.4 is not released: LLaMA 3.3:70B https://www.llama.com/llama-downloads/

      Attachments

        Activity

          Loading...
          Uploaded image for project: 'translate5'
          1. translate5
          2. TRANSLATE-4438

          Integrate Llama 3 as language resource analogous to GPT

          Details

            • Improvement
            • Resolution: Unresolved
            • None
            • None
            • openai

            Description

              Requirement

              Llama should be integrated in translate5 as trainable language resource completely analogous to what we are doing at the moment with OpenAI and Azure GPT.

              • persistent training of models should be possible with prompts and exmaples
              • The new training lib features as developed in https://jira.translate5.net/browse/TRANSLATE-4369 should be possible to be used
              • Context matters. We should send as many segments as possible at the same time to Llama. We should count the tokens before sending the request for translation to Llama, to make sure, we send as many segments as possible for translation.
              • We should send the terminology to Llama, in the way, we do for GPT and be prepared, that we will implement also https://jira.translate5.net/browse/TRANSLATE-4401 and its sub-issues soon for GPT and for Llama
              • We should evaluate, if there is  a better RAG-based approach for enriching GPT with the terminology (and soon matches from the TM) than what we are currently doing
              • We should have in mind, that for quality estimation we will probably also need and use the training interfaces

              Llama model to be used as long as Llama 3.4 is not released: LLaMA 3.3:70B https://www.llama.com/llama-downloads/

              Attachments

                Activity

                  People

                    leonkiz Leon Kiz
                    marcmittag Marc Mittag [Administrator]
                    Aleksandar Mitrev
                    Votes:
                    0 Vote for this issue
                    Watchers:
                    1 Start watching this issue

                    Dates

                      Created:
                      Updated:

                      People

                        leonkiz Leon Kiz
                        marcmittag Marc Mittag [Administrator]
                        Aleksandar Mitrev
                        Votes:
                        0 Vote for this issue
                        Watchers:
                        1 Start watching this issue

                        Dates

                          Created:
                          Updated: