Indian Flag
Government Of India
A-
A
A+
/IITGN ORGANISATION
UpvoteUpvote
1

Ganga-2-1B

The first pre-trained Hindi model by any academic research lab in India 🇮🇳!
  • Upvotes1upvotes
  • Downloads50downloads
  • Views689views
  • File size1.88 GB
  • Shares0Shares
  • Last updatedupdated 31 Jul 2025
UpvoteUpvote
1
LICENSE Apache 2.0
SECTOR Science, Technology and Research

About Model

Project Unity is an initiative to address India's linguistic diversity and richness by creating a comprehensive resource covering the country's major languages. We strive to achieve state-of-the-art performance in understanding and generating text in Indian languages. To achieve this, we train models on the monolingual regional languages of India. Our first release is the Ganga-1B model, which has been trained on a large dataset of public domain web-crawled Hindi language data, including news articles, web documents, books, government publications, educational materials, and social media conversations (filtered for quality). Additionally, the dataset has been further curated by native Indian speakers to ensure high quality. Significantly, the Ganga-2-1B model outperforms existing open-source models that support Indian languages, even at sizes of up to 7 billion parameters.

Developed by: Lingo Research Group at IIT Gandhinagar

Model type: Autoregressive Language Model

Language(s): Bilingual (Primary: Hindi [hi], Secondary: English [en])

The instruct tunned model Ganga-2-1b trained on a monolingual Hindi language dataset as part of Project Unity. We propose the name Ganga 🌊 to honor the longest river flowing through the Hindi-speaking region of India 🇮🇳.

Disclaimer: This is a text-completion model designed for fine-tuning on downstream tasks. It is not intended for direct use as a chat or instruction-following model.

Technical Specifications 🤖
    • Precision: BFloat16
    • Context Length: 2,048
    • Learning Rate: 4e-4
    • Optimizer: AdamW
    • LR Scheduler: Cosine

    Model Architecture and Objective: 

    Ganga-2-1b is a decoder-only transformer model featuring the following specifications:

      • Layers: 16
      • Attention heads: 32
      • Embedding dimension: 2,048
      • Vocabulary size: 32,768
      • Sliding window: 1024
      • Intermediate dimension: 7,168

      Results:

      Tokenizers Results:

      ModelFertility
      Ganga-2-1b1.12
      Pragna-1b1.58
      Bloom-1b11.27
      Bloom-1b71.27
      Gemma-2b1.89
      Bloom-3b1.27
      Airavata-7b1.69
      Sarvam-2b1.38

      Metrics:
      ModelPPLSangraha Dataset
      Ganga-2-1b8.09
      Ganga-1b15.82
      Pragna-1b9.37
      Bloom-1b117.49
      Bloom-1b714.28
      Gemma-2b31.01
      Bloom-3b12.82
      OpenHathi-7B25.73
      Airavata-7b38.24
      Sarvam-2b10.31

      Recommendations ‼️

      This model described is a research preview and is under ongoing iterative updations, and as such, it only provides limited safety measures. Additionally, it may generate offensive content. It is strictly prohibited to use the model for any illegal, harmful, violent, racist, or sexual purposes.

      Metadata Metadata

      Apache 2.0

      Science, Technology and Research

      Aamod Thakur, Mayank singh

      N.A.

      IITGN

      N.A.

      N.A.

      N.A.

      N.A.

      N.A.

      N.A.

      N.A.

      Open

      20/08/25 05:43:32

      N.A.

      1.88 GB

      Tags

      • Text Generation

      Related Models by Organization Related Models by Organization

      COMI-LINGUA-POS
      This is a fine-tuned version of aya-expanse-8b for Part-of-Speech (POS) Tagging on Hinglish (Hindi-English code-mixed) text. It assigns a grammatical category to each token using a language-agnostic Universal POS tagset suitable for code-mixed content in Roman and Devanagari scripts.
      IITGN
      • See Upvoters0
      • Downloads6
      • Views180
      • File Size979.67 MB

      COMI-LINGUA-MT
      This is a fine-tuned version of Llama-3.1-8B-Instruct for Machine Translation (MT) on Hinglish (Hindi-English code-mixed) text. It translates code-mixed input in Roman/Devanagari scripts to three target formats: (i) Standard English, (ii) Romanized Hindi, and (iii) Devanagari Hindi.
      IITGN
      • See Upvoters0
      • Downloads9
      • Views229
      • File Size1.89 GB

      COMI-LINGUA-MLI
      This is a fine-tuned version of aya-expanse-8b for Part-of-Speech (POS) Tagging on Hinglish (Hindi-English code-mixed) text. It classifies each sentence at the sentence level into the dominant matrix language governing the grammatical structure: hi (Hindi) or en (English).
      IITGN
      • See Upvoters0
      • Downloads6
      • Views203
      • File Size1.89 GB

      COMI-LINGUA-LID
      This is a fine-tuned version of aya-expanse-8b for Token-level Language Identification (LID) on Hinglish (Hindi-English code-mixed) text. It performs token-wise classification into three categories: en (English), hi (Hindi), or ot (Other).
      IITGN
      • See Upvoters0
      • Downloads17
      • Views263
      • File Size1.89 GB

      COMI-LINGUA-NER
      This is a fine-tuned version of aya-expanse-8b for Named Entity Recognition (NER) on Hinglish (Hindi-English code-mixed) text. It helps with token-level entity tagging (PERSON, ORGANISATION, LOCATION, DATE, TIME, GPE, HASHTAG, EMOJI, MENTION, X/Other) in Roman/Devanagari scripts. Achieves 94.90 F1 on COMI-LINGUA test set (5K instances), outperforming the zero-shot inference (59.88 F1).
      IITGN
      • See Upvoters0
      • Downloads10
      • Views293
      • File Size979.66 MB


      Related Models by Sector Related Models by Sector

      shuka-v1
      Multilingual audio to text model
      SARVAM AI
      • See Upvoters3
      • Downloads117
      • Views2,677
      • File Size0

      sarvam-1
      India's first indic model, pretrained on 4 trillion tokens
      SARVAM AI
      • See Upvoters4
      • Downloads305
      • Views6,837
      • File Size0

      LUCAN 10B
      LUCAN 10B Physics MoE is an 11B parameter multi-domain scientific foundation model built on a hybrid Mamba2 state-space and 12-layer Sparse Mixture-of-Experts (MoE) architecture with a PhysicsConstrainedOutputHead. It models continuous scientific latents across fluid transport, chemical reactors, phase-change thermal physics, and molecular systems.
      SHODH AI PRIVATE LIMITED
      • See Upvoters0
      • Downloads0
      • Views38
      • File Size4.02 GB

      LUCAN 3B
      LUCAN 3B is an approximately 3-billion-parameter scientific AI model from Shodh AI for processing numerical reactor data. This release includes the model weights, dedicated tokenizers and matching adapters for bioreactor and continuous-flow reactor applications, along with loading instructions and an introductory guide.
      SHODH AI PRIVATE LIMITED
      • See Upvoters0
      • Downloads1
      • Views43
      • File Size4.44 GB

      MIZO MT
      This Machine Translation (MT) model translates text from one language to another using a trained neural network. It can support multilingual communication and language translation applications. Translation quality depends on language coverage, training data, and contextual accuracy. For Mizo, evaluate the model using representative test sentences before deployment.
      DIGITAL INDIA BHASHINI DIVISION
      • See Upvoters0
      • Downloads0
      • Views23
      • File Size496.68 MB

      COILD Multilingual Translation
      COILD Multilingual Translation Model (COILD-Mul-MT)
      DIGITAL INDIA BHASHINI DIVISION
      • See Upvoters0
      • Downloads0
      • Views36
      • File Size0

      GNM Generative aNthropometric Model - v3
      GNM (pronounced genome) is a state-of-the-art parametric 3D statistical model of the human head developed by Google. Learned from a large dataset of high-resolution 3D scans, GNM provides fine-grained, disentangled control over facial identity, expressions, and head pose, complete with controllable internal anatomy (eyeballs, teeth, and tongue). The model is released under the Apache 2.0 permissive license, suitable for both academic research and commercial applications.
      GOOGLE LLC
      • See Upvoters0
      • Downloads0
      • Views33
      • File Size0

      BHASHINI IISc Sourashtra VITS TTS Models
      VITS text-to-speech models for Sourashtra, a low-resource Indo-Aryan language. There are male and female voices, and each voice accepts text in either Tamil script or Sourashtra script.
      DIGITAL INDIA BHASHINI DIVISION
      • See Upvoters1
      • Downloads3
      • Views117
      • File Size0

      DiffusionGemma
      DiffusionGemma is a generative model built by Google DeepMind. Based on the 26B A4B Mixture-of-Experts (MoE) Gemma 4 architecture, DiffusionGemma generates tokens using discrete diffusion. This open-weights model is multimodal, handling text, image, and video inputs to generate text output.
      GOOGLE LLC
      • See Upvoters1
      • Downloads4
      • Views70
      • File Size0

      SPRING-INX-DATA2VEC-AQC-GUJARATI
      Automatic Speech Recognition (ASR) model for speech recognition, processing audio and transcribing spoken content into text.The inference code, installation requirements, and usage instructions are available in the SPRING Lab, IIT Madras GitHub repository: https://github.com/Speech-Lab-IITM/Fairseq-Inference
      DIGITAL INDIA BHASHINI DIVISION
      • See Upvoters0
      • Downloads4
      • Views64
      • File Size3.52 GB


      Related Models by Task Type Related Models by Task Type

      Dhir-5B-4bit
      Dhīr 5B MoE Q4_K_M GGUF (3.4 GB) for offline llama.cpp in English + 10 Indian languages. 5.09B total/1.79B active parameters; 91% token agreement (1.5M tokens), ≤1.8% perplexity vs BF16.
      Large Language Models
      ZENTEIQ
      • See Upvoters3
      • Downloads16
      • Views221
      • File Size3.13 GB

      Dhir-2B-8bit
      Dhīr 2B Q8_0 GGUF (2.2 GB) for offline llama.cpp. Versus BF16: WikiText-2 perplexity within 0.1%, with 98% next-token agreement over 1.5M tokens. English + 10 Indian languages.
      Large Language Models
      ZENTEIQ
      • See Upvoters0
      • Downloads7
      • Views97
      • File Size2.02 GB

      Dhir-2B-4bit
      Dhīr 2B quantised to 4-bit GGUF for llama.cpp. At just 1.3 GB, it runs fully offline on mobiles, embedded chips and laptops in English and ten Indian languages, without requiring a GPU.
      Large Language Models
      ZENTEIQ
      • See Upvoters5
      • Downloads27
      • Views181
      • File Size1.20 GB

      Chitrarth - Multilingual Vision-Language Model
      Chitrarth is a multilingual vision-language model (VLM) integrating a Large Language Model (LLM) with a vision module. It is trained on multilingual image-text data and supports 10 Indic languages along with English.
      OLA KRUTRIM
      • See Upvoters1
      • Downloads145
      • Views2,389
      • File Size0

      Krutrim-1 Instruct Large Language Model
      Krutrim-1 is a 7.3B parameter multilingual foundation model trained on a 2 trillion token dataset, designed for Indian linguistic and demographic needs. It supports 11 Indic languages and matches or exceeds comparable state-of-the-art models in multilingual tasks.
      OLA KRUTRIM
      • See Upvoters0
      • Downloads28
      • Views1,381
      • File Size0

      Vyakyarth - Multilingual Sentence Embedding Model
      Vyakyarth is a sentence-transformers-based model fine-tuned for Indic languages, capable of mapping text to a 768-dimensional dense vector space for semantic search, similarity, classification, and clustering tasks.
      OLA KRUTRIM
      • See Upvoters1
      • Downloads53
      • Views1,244
      • File Size0

      Krutrim-2 Instruct
      Krutrim-2 is a 12B parameter multilingual large language model built on the Mistral-NeMo 12B architecture, optimized for Indic languages and Indian cultural context. It supports long-form conversations, reasoning, coding, and translation tasks.
      OLA KRUTRIM
      • See Upvoters0
      • Downloads33
      • Views1,114
      • File Size0

      Dhir-5B
      Dhīr 5B MoE is a compact 5.09-billion-parameter Mixture-of-Experts foundation model developed and pretrained from scratch by ZenteiQ. About 1.79 billion parameters are active for each token. With support for English and ten Indian languages, it is trained for instruction following, conversational reasoning, coding, structured output, document-grounded workflows and tool use, and is designed for private, local and on-premises AI applications.
      ZENTEIQ
      • See Upvoters11
      • Downloads100
      • Views1,854
      • File Size9.50 GB

      Dhir-2B
      A compact 2.03B-parameter foundation model, pretrained from scratch in India, for private and edge AI in English and ten Indian languages.
      ZENTEIQ
      • See Upvoters15
      • Downloads117
      • Views4,623
      • File Size3.79 GB

      YoosaHealthAI
      Yoosa Health AI is a revolutionary medical intelligence engine by Webdpro Solutions. Built on Mistral 7B, it serves patients, doctors, and hospitals via four core modules: Detect, Diagnose, Talk (Voice AI), and Suggest. It streamlines clinical workflows through report summarization and automated triaging while boosting patient literacy. Targeting a 1B parameter model by 2027, Yoosa aims to shift global healthcare from reactive sick care to proactive, accessible, and scalable intelligence.
      WEBDPRO SOLUTIONS PRIVATE LIMITED
      • See Upvoters0
      • Downloads4
      • Views129
      • File Size0