The Guide to Building Open Indic LLMs Today

[ramsrigouthamg]

Title of the talk/workshop
The Guide to Building Open Indic LLMs Today

Abstract of the talk/workshop

• Steps in training modern LLMs
• Challenges specific to Indic Models (Tokenizer + Instruction and Evaluation Datasets)
• Steps in training an Indic LLM
• Techniques that make this possible today on single consumer grade GPU (Quantization, LORA etc)
• Frameworks that further enable speed of training (Unsloth etc)
• Best LLMs suited for Fine-tuning Indic LLMs (Meta Llama 2/3, Google’s Gemma etc)
• How we trained Navarasa 2.0 catering to 15 Indian Languages based on Google’s Gemma.
• Where do we need to make progress? (Regional Datasets for SFT and evaluation, Fine Tuning techniques like ORPO)
• Evolving training techniques (DORA etc)

Category of the talk/workshop
Data Science, Machine Learning, and AI

Duration (including Q&A)
45 mins

Level of Audience
Beginner/Intermediate

Speaker Bio
Ramsri is an open-source developer of Indic finetuned LLMs as well as builder of AI SaaS apps (Questgen.ai and Supermeme.ai) from idea to 600k+ users.
Email: [email protected]
Years of Experience: 11 yrs

Prerequisites(if any)
Basic knowledge of AI and ML

[kalyan678]

@bhansa - Checkout this details and create flyer accordingly. You can use their SM's to grab the picture :)