Mlops
CurrentWon "SMM4H-2024 Task 2 – Cross-Lingual Few-Shot Relation Extraction for Pharmacovigilance in French, German, and Japanese"SageMaker Pipeline Evaluation step using lm-evaluation-harness on single GPU (g4dn machine) with help of CPU offloadingBuilt TGIv1.3.4 container for sagemaker to deploy Mistralv0.2LLM Token Counter on AWS Lambda-- improved inference 8x by not using transformers libraryBuilt TGIv1.0.0 container for sagemaker to deploy Llama2-70b-chatpython boto3 workflow to install a Customized Groundtruth job for multilabel annotation of text pairs. It included:-- setting up necessary Cognito infrastructure-- deploying preprocess and postprocess Lambda for the annotation job-- generating appropriate IAM policyLoad Testing Sagemaker Endpoint using:-- locust.io-- intuit's sagemaker-gatling (with modified AWS Signature Version 4)-- Inference Recommender-- K6 on AWS Lambda (https://github.com/philschmid/text-generation-inference-tests)AWS Lambda to evaluate a LLM output using CER (Character error rate)Sagemaker Project to deploy fine-tuned Dolly-v2-7b on hugginface TGI containerSupervised Fine-tuning of Dolly-v2-7b using huggingface TRL- Visualization tool for Dolly-v2 LLM--- displays attention (based on Jesse Vig's Bertviz)--- displays transition scoresPrototype Annotation / Data Management Layer on top of argilla / dvc / prodigy
