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 – 150 Hours Training Test Results | Base Model Output Analysis

Overview

After 150+ hours of training, HTGM v2 Hindi LLM was tested using real Hindi question-answer prompts to evaluate how the base model is learning.

This is still the base model stage — not fine-tuned yet — and only around 4% of total planned training has been completed.

The goal of this test was simple:

To check whether the model is learning factual patterns, language structure, and question-answer understanding.

This repository documents the real outputs, observations, and current model status.


Project Details

  • Model Name: HTGM v2
  • Type: Hindi Large Language Model (Hindi LLM)
  • Architecture: GPT-style Transformer
  • Parameters: 163.4M+
  • Sequence Length: 2048
  • Vocabulary Size: 100,000
  • Dataset Size: ~41GB
  • Dataset Source: AI4Bharat Sangrah + Manual Cleaning
  • Training Platform: Kaggle
  • GPU: 2x T4
  • Tokenizer: Hugging Face BPE-based Tokenizer

Training Status

Current Progress

  • Training Started: 1 April
  • Total Training Time: 150+ Hours
  • Total Days: 24 Days
  • Sessions Completed: 17
  • Optimizer Steps: 38K+
  • Total Training Completion: ~4%
  • Training Still Running

This means the model is still in an early-stage pretraining phase.

A lot more learning is still left.


Model Testing Setup

Inference Configuration

  • Model Loader: GPT2LMHeadModel
  • Device: CPU
  • Tokenizer: Custom BPE Tokenizer

Generation Settings

  • temperature: 0.7
  • top_k: 50
  • top_p: 0.9
  • repetition_penalty: 1.2

These settings were used to reduce repetition and improve generation stability.


Real Test Outputs

These are real outputs directly generated by HTGM v2.

No manual editing.


Test Example 1

Input

प्रश्न: हिंदी भाषा किस देश में बोली जाती है? उत्तर:

Result

The model understood the topic was related to language and words, but failed to directly answer “भारत”.

It generated broken structure and drifted into unrelated patterns.

Observation

Pattern recognition exists, but factual answering is still weak.


Test Example 2

Input

प्रश्न: सूर्य किस दिशा से निकलता है? उत्तर:

Result

The model understood the concept of direction and time, but failed to correctly answer “पूर्व दिशा”.

Instead, it generated repetitive reasoning.

Observation

Context recognition exists, but answer completion is unreliable.


Test Example 3

Input

प्रश्न: भारत की राजधानी क्या है? उत्तर:

Result

The model correctly started with:

दिल्ली

…but then drifted into unrelated repetitive text.

Observation

This is a strong sign that factual learning has started.

This is an important milestone.


Test Example 4

Input

प्रश्न: तुम कैसे हो? उत्तर:

Result

The model understood conversational format but entered heavy repetition like:

“बस बस बस…”

Observation

Conversation structure is learned, but coherence is still weak.


What These Results Mean

This is not failure.

This is exactly how base model learning works.

At this stage, success means:

  • Correct starting tokens
  • Pattern recognition
  • Q&A format understanding
  • Partial factual recall

HTGM v2 is already showing these signs.

That is real progress.


Current Limitations

  • High repetition
  • Weak coherence
  • Context drifting
  • Incomplete factual answers
  • No instruction fine-tuning yet

This is expected because the model is still only around 4% trained.


Next Steps

Planned Improvements

  • More pretraining
  • Supervised Fine-Tuning (SFT)
  • Better dataset structuring
  • Stronger instruction-following
  • Reduced repetition
  • Better output alignment

This is where major quality improvements will happen.


Conclusion

After 150+ hours, HTGM v2 is not perfect.

But it is learning.

It recognizes patterns.
It starts correct answers.
It understands structure.
It shows real progress.

This is still the beginning — not the final model.

But the foundation is becoming real.


Author

Mahesh Editor
India AI Official


Build in Public

No fake hype.
Only real outputs.
Only real progress.

This is how real AI gets built.

About

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

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors