hindi-llm

Rajbhasha.net is a pioneering Hindi computing platform dedicated to simplifying Indian language technology for everyone. The project focuses on Hindi typing, Unicode conversion, font utilities, OCR, voice typing, and official-language workflows widely used across government offices, educational institutions, publishers, and digital creators.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Session 19 crosses 165+ hours training and 42K+ optimizer steps milestone.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, checkpoints, experiments, and progress shared publicly.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, checkpoints, experiments, and progress shared publicly.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, checkpoints, experiments, and progress shared publicly.

Real test results of HTGM v2 Hindi LLM after 150+ training hours. Base model output analysis, Hindi Q&A testing, and learning progress.

HTGM v2 Hindi LLM reaches 180+ hours and 48K+ steps. Session 21 shows high learning rate instability and real LLM training dynamics.

HTGM v2 Hindi LLM trained for 220+ hours on 41GB dataset. Real model outputs, limitations, and insights from base pre-training before SFT.

HTGM v2 Hindi LLM reaches 210+ hours and 54K+ steps. Session 23 shows stable validation despite high training fluctuations and real LLM training dynamics.

HTGM v2 crosses 150+ training hours while Mahesh Editor starts Smell AI, a new AI research project focused on smell and human emotional reactions.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Session 18 crosses 155+ hours training and 40K+ optimizer steps milestone.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, checkpoints, experiments, and progress shared publicly.

HTGM v2 Hindi LLM reaches 170+ training hours and 46K+ steps. Real training logs, fluctuation phase, and GPT-based Hindi AI development journey.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, checkpoints, experiments, and progress shared publicly.

HTGM v2 Hindi LLM completes 220+ training hours and 56K+ steps in Session 24. Validation loss improves as pre-training nears completion before SFT.

HTGM v2 Hindi LLM completes 230+ training hours and 59K+ optimizer steps. Session 25 shows improved validation loss, lower perplexity, and stable convergence.

HTGM v2 Hindi LLM crosses 200+ hours and 51K+ steps. Session 22 shows recovery from instability with improved validation and stable training pipeline.

Building HTGM v2 — a Hindi LLM from scratch using GPT architecture. Real training logs, experiments, and progress shared in public.

HTGM.2 is a Hindi-first BPE tokenizer trained on ~41GB corpus using streaming architecture for scalable Hindi LLMs, Devanagari NLP, and low-memory tokenizer engineering.

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.