SM Challenges

PEFT Methods

Additive
Fine-Tunning

Adapter-based

Adapter Design

Multi-task Adaptation

Soft Prompt-based

Soft Prompt Design

Training Speedup

Reparameterized
Fine-Tuning

Low-rank
Decomposition

LoRA Derivatives

Dynamic Rank

LoRA Improvement

Multiple LoRA

Selective
Fine-Tunning

Unstructural Masking

Structural Masking

LLMs Tasks

Language Modeling

Text Generation

DataOps

Data Pipeline Orchestration
Managing the flow from data ingestion
to storage and processing.

Data Sorces

1. Base data: User, Expert, ...tabular , text , geospatial

Product

2. Product Meta , Details, Categories

3. User Reviews

4. User clickstream data, system log, user preference

5. QA interactions (Bellysafe, Expert, ...) Text-based data.

6. Calendar Tracking: Time series data (appointments, events, etc.).

7. External, Knowledge_Base

Ingestion Layer

Batch ingestion: large data dumps like historical product data or user logs

Real-time ingestion: For handling QA, user chats , calendar event updates

Data Storage / Data Lake

AWS S3: suitable for building a data lake.

OpenSearch: If search and retrieval are key functions, OpenSearch (for analytics) can be used in combination with a data lake.

Version Control & Collaboration: Managing different versions of data, changes in ML models, and collaboration among team members.

Monitoring and Quality Assurance: Ensure data quality, detect anomalies, and apply corrective actions.

Security & Compliance: Ensuring data privacy (important for health-related apps) and compliance with regulations like HIPAA.

Product Use Cases

Product Similarity/relevancy

Rrecommendation

Category Rec

Popularity_based

Personalized

Product Basket

Customers Say

Summarization

Sentiment Analysis

Attributes/Tag generation

AI Criteria Tags

Search + Result Page

Filter + Result Page

Search

Semantic

FullText