Skip to content

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

HTGM v2 Hindi LLM – Pre-Training Near Completion 🚀

After 220+ hours of training across 24 sessions, HTGM v2 Hindi LLM is now approaching the end of its base pre-training phase.

This is not an official final stop yet, but training may be paused to avoid overfitting. If needed, training can still continue further.


HTGM Journey

HTGM v1

  • ~9GB Hindi dataset
  • Experiment phase
  • Learning training pipeline

HTGM v2

  • ~41GB Hindi dataset
  • Foundation phase
  • 220+ hours training
  • ~163M parameter GPT-style model

Current Focus

Now the next major step is:

SFT (Supervised Fine-Tuning)

The upcoming work will focus on:

  • QA datasets
  • Instruction-following
  • Better response formatting
  • Improved reasoning
  • Higher answer accuracy

Future Vision – HTGM v3

HTGM v3 is currently a future research vision.

Target:

  • 1TB → 10TB high-quality Hindi dataset

This would be a massive scale jump compared to:

  • HTGM v1 → 9GB
  • HTGM v2 → 41GB

Future Compute Requirements

Kaggle GPUs were enough for experimentation and learning.

But future HTGM v3 training may require:

  • NVIDIA A100 GPUs
  • NVIDIA H100 GPUs
  • Enterprise-level infrastructure
  • Multi-GPU systems

Support from startups, companies, or research organizations may be needed in the future.


Important Reality Check

HTGM v3 is NOT starting now.

There is still a lot of work left:

  • Better datasets
  • Better optimization
  • Better fine-tuning
  • Better architecture

Right now, the focus remains on improving HTGM v2.


Final Thought

HTGM v1 → Experiment
HTGM v2 → Foundation
HTGM v3 → Scale & Intelligence

Small steps today. Bigger AI tomorrow.

— Mahesh Editor

About

HTGM v2 Hindi LLM crosses 220+ training hours and approaches pre-training completion. Future plans include SFT training and HTGM v3 with TB-scale Hindi datasets.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors