Training Improvements Master Ticket

[sooraj1002]

Frontend Dependencies

• Share GET for datasets

Backend

• API for initiating training should accept Metadata to override default configurations in specific instances #87
• Create API to get metadata to fill model card - this will be workflow agnostic. Will only take the userId and the task type
• How to publish model to places other than Huggingface #89
• How to deploy a model once its at huggingface  #84 @TakshPanchal - TBD by 12/06/24
• API to see scheduled training tasks - Expose the celery queue
• API to see all models which have been trained - expose model registry
• Monitor GPU usage for a task #85 - @singhalkarun to provide an API to share task wise status of GPU usage
• update training API to allow for scheduled training of model - Schedule it in celery
• Setup Initial APIs for Model Management Page for Deployment #86
• Sagemaker #88
• RLHF Pipeline Requirements #96

[KDwevedi]

NVIDIA Triton: https://github.com/triton-inference-server/server

[KDwevedi]

Sooraj

1. Deployment
2. Model Read

Karan

1. Datasets
2. Model - [Name] [Task] [Dataset] [Last Tuned] [Last Deployed] [Deploy]
3. Training Datasets - [Name] [Desc] [Model] [Taks] [Last Update]

A flow for generating relevant examples and closing those

[KDwevedi]

[ASR] (iterate) -> [Transliteration] -> [Translation] -> [Intent Classification] -> [NER] -> [T2S]

[KDwevedi]

• E2E for a model Label Studio, Dataset, Model, Deployment Button
• Generating training samples
  • Label Studio Feedback
    • Ask Autotune to generate 5-10 training examples for a given correction
    • Prompt, Config will have to be part of model config --> workflow
• Video Demo

How to plant wheat?

text_classification
--> Computer Programming
--> Agriculture

How to plant beets?
What is wheat about?
--> Agriculture
--> Agriculture

FE
1,2 -> Correct
3,4,5 -> Discard

[KDwevedi]

closing this ticket as not required, all remaining contextual tasks have been moved to #117, #118