Khasi language model trained on 3.6M sentences using RoBERTa architecture. 110M parameters. Supports NLP tasks for Khasi text processing.
1upvotes
33downloads
1,314views
0
0Shares
updated 15 Sept 2025
Upvote
1
LICENSE Creative Commons Attribution Non Commercial 4.0
SECTOR Education and Skill Development
About Model
KhasiBERT is the first transformer-based language model developed specifically for the Khasi language of Meghalaya. This foundational model addresses the computational linguistics needs for Meghalaya's 1.4+ million Khasi speakers and enables digital language processing infrastructure for the state.
The model implements the RoBERTa architecture with 12 transformer layers, 768 hidden dimensions, and 12 attention heads, totaling 110,652,416 parameters. KhasiBERT was trained using masked language modeling on a corpus of 3,621,116 Khasi sentences collected and preprocessed to represent diverse linguistic patterns of the language.
A custom Byte-Level BPE tokenizer with 32,000 vocabulary tokens was developed specifically for the Khasi language to handle its Austroasiatic linguistic characteristics effectively. Training utilized mixed-precision (FP16) optimization with a batch size of 24, learning rate of 5e-5, and AdamW optimizer over 150,880 training steps.
The model supports standard transformer operations including masked token prediction, contextualized embeddings generation, and can be fine-tuned for downstream tasks such as text classification, sentiment analysis, named entity recognition, and question answering in Khasi. Input sequences are limited to 512 tokens with standard special tokens for beginning-of-sequence, end-of-sequence, padding, unknown tokens, and masking.
Technical specifications include GELU activation functions, 0.1 dropout probability, layer normalization with epsilon 1e-12, and positional embeddings up to 514 positions. The model weights are distributed in Safetensors format and are compatible with the Transformers library ecosystem for integration with existing NLP workflows.
Northeast India's first AI language model. Kren-M is a 2.6B parameter bilingual model for Khasi-English, built on Gemma-2-2B. Features Kren-NE custom tokenizer covering 7 NE languages (Khasi, Garo, Mizo, Assamese, Manipuri, Nagamese, Nyishi) with 35.7% efficiency gain. Trained on 5.43M Khasi sentences. Capabilities: bidirectional translation, natural conversation, cultural context. Designed for language preservation across Northeast India
A multilingual Speech-to-Text (STT) model for eight Northeast Indian languages, fine-tuned from Whisper Medium using over 150,000 speech-text pairs from public and institutional datasets. The model expands speech recognition support for low-resource indigenous languages, including Khasi, Garo, Mizo, Kokborok, Nagamese, Assamese, Chakma, and Wancho.
NE-SpeechEmbed is a multilingual speech-text embedding model by MWire Labs for Northeast Indian languages. The model supports semantic speech search, cross-modal retrieval, and audio-text embeddings across Khasi, Garo, Mizo, Nagamese, Kokborok, Assamese, Wancho, and Chakma.
NE-Embed is a multilingual text embedding model for Northeast Indian languages, enabling semantic search, retrieval, and RAG across 10 languages including Khasi, Garo, Meitei, Bodo, Mizo, Assamese, Nyishi, Kokborok, Pnar, and Nagamese. Fine-tuned on LaBSE with 201,738 parallel pairs.
NE-OCR is a multilingual Optical Character Recognition model developed by MWire Labs to accurately recognize printed text from documents in Northeast Indian languages. The model supports Assamese, Bodo, English, Garo, Hindi, Khasi, Kokborok, Meitei (Bengali script), Meitei (Meitei Mayek script), Mizo, Nagamese, and Nyishi. It is designed to enable reliable digitization of books, newspapers, government records, educational materials, and cultural archives from Northeast India where mainstream OCR
NE-LID is a fast and accurate language identification model for Northeast Indian languages using character level features. It is designed for low resource and script diverse text and achieves high accuracy on short sentences.
NortheastNER is a token classification model built on XLM-RoBERTa and fine-tuned on ~25k sentences from gazetteers, news, and cultural texts across Northeast India. It detects region-specific entities, places, tribes, festivals, tourist sites, flora, fauna, and experimental local names; ideal for low-resource NER, regional search, cultural analytics, and knowledge graph applications.
Northeast India's first AI language model. Kren-M is a 2.6B parameter bilingual model for Khasi-English, built on Gemma-2-2B. Features Kren-NE custom tokenizer covering 7 NE languages (Khasi, Garo, Mizo, Assamese, Manipuri, Nagamese, Nyishi) with 35.7% efficiency gain. Trained on 5.43M Khasi sentences. Capabilities: bidirectional translation, natural conversation, cultural context. Designed for language preservation across Northeast India
IndNumNLI Classifier is a soft-voting ensemble of Random Forest and Gradient Boosting trained on 16 numerical reasoning features extracted from Indian government statistical premise-hypothesis pairs. It classifies NLI labels as entailment, neutral, or contradiction with a test accuracy of 60 percent and macro F1 of 0.64.
NEET-BioBERT is a fine-tuned version of DistilBERT (base uncased) specifically trained to classify the correct option for NEET-style multiple-choice biology questions. It selects the best answer among four choices (A, B, C, D).
A multimodal Transformer model designed for generating grounded and non-grounded radiology reports from chest X-rays, integrating vision and language understanding for medical AI research.
BharatGPT-Instruct E2B is a multilingual language model by CoRover.ai, designed to deliver high-quality, culturally relevant AI responses across Indian languages.
BharatGPT Mini - A compact, Transformer-based language model developed by CoRover, pretrained on large-scale text data to understand and generate natural language, advancing India’s indigenous AI capabilities.
A multilingual model fine-tuned on authentic conversational data across Indic languages, supporting translation, summarisation, and text generation. Ideal for Secure RAG (S-RAG) and domain-specific fine-tuning, it enables the development of sovereign, multilingual AI agents, chatbots, voicebots, and virtual assistants.
Kaveri-4B is a multilingual text-generation model for conversational AI, instruction following, reasoning, coding assistance, summarization, translation, and general-purpose NLP tasks.
Kaveri-1.5B is a lightweight multilingual text-generation assistant by RiyaVibe for general chat, instruction following, basic question answering, simple coding help, basic math, translation, and local or cloud inference.
A 1.3B parameter Transformer model trained on structured QA, NLP tasks, and Python code, optimized for text generation, summarization, and creative writing, with no fine-tuning from human feedback.