Official Research Portal powered by - Centre for Innovation, Incubation and Entrepreneurship (CIIE), Moran College, Assam, India
অসমীয়া ভাষা প্ৰযুক্তি আৰু কৃত্ৰিম বুদ্ধিমত্তা গৱেষণাগাৰ
INDIGENOUS AI RESEARCH INITIATIVE · কৃত্ৰিম বুদ্ধিমত্তা গৱেষণা পদক্ষেপ

Indigenous Intelligence. Global Innovation.

AssamAI Lab is an artificial intelligence research and development initiative focused on Assamese and underrepresented languages. We develop language models, explore emerging AI technologies, and build accessible solutions for multilingual communities.

AssamAI Lab is an artificial intelligence research and development initiative of the Centre for Innovation, Incubation and Entrepreneurship (CIIE), Moran College, Assam, India.
Latest AI Models · Active Releases

Indigenous Assamese Language Model Suite

AssameseLM-E2B Available

Foundation causal language model pretrained on normalized Assamese Unicode corpora (অসমীয়া) for base representation & fine-tuning.

Base LM · E2B
AssameseLM-E2B-it Available

Instruction-tuned Assamese variant aligned for Q&A, essay & application writing, summarisation, paraphrasing, and reasoning.

Instruction SFT
Spinning Indigenous Heritage Mandala
অAI
Indigenous Heritage × Sovereign AI
Rooted in Assamese Linguistic & Cultural Identity

Bridging traditional regional knowledge systems with modern transformer architectures at AssamAI Lab.

Research Architecture

Assamese Language Modelling & Adaptation Pipeline

E2B Series
STAGE 01 · CORPUS & TOKENIZATION অসমীয়া পাঠ্য সংগ্ৰহ

Assamese Unicode normalization (U+0980–U+09FF), morphological segmentation, and tokenizer vocabulary adaptation for low-resource script efficiency.

↓ Continued Pretraining & Domain Adaptation
STAGE 02 · FOUNDATION MODEL Available
AssameseLM-E2B

Base Assamese language model focused on linguistic representation, text completion, and cross-lingual transfer.

↓ Supervised Instruction Tuning (SFT)
STAGE 03 · INSTRUCTION-TUNED MODEL Available
AssameseLM-E2B-it

Instruction-tuned variant for Assamese question answering, Retrieval-Augmented Generation (RAG), summarisation, essay/application writing, and reasoning.

↓ Evaluation & Downstream Research
STAGE 04 · MULTILINGUAL ECOSYSTEM PPL · BPB · Factuality

Translation, ASR, TTS, Assamese OCR, Document Intelligence, and locally deployable educational assistants.

Laboratory Mandate

Intelligence Rooted in Language. Innovation Beyond Boundaries.

AssamAI Lab works at the intersection of artificial intelligence, language technology, and research. Our work focuses on developing language models and practical AI systems that address the needs of Assamese and other underrepresented languages.

Rather than operating as a commercial startup, AssamAI Lab is structured as an academic-anchored research and development platform dedicated to scientific integrity, reproducible evaluation, and locally deployable language technologies that strengthen India's multilingual AI ecosystem.

Institutional Structure & Governance
01
Moran College
Host Academic Institution · Moranhat, Assam, India
↳ Institutional Parent
02
Centre for Innovation, Incubation and Entrepreneurship (CIIE)
Research Incubation & Innovation Body, Moran College
↳ Dedicated Research Initiative
03
AssamAI Lab
Indigenous AI Research, Language Models, Technologies & Applications
Model Catalogue

Featured Language Models & Research Roadmap

Clear distinction between released models available for evaluation and planned research directions.

FOUNDATION LANGUAGE MODEL
Available

AssameseLM-E2B

Base Assamese language model engineered for indigenous text representation, continued pretraining research, Assamese Unicode sequence modelling, and downstream task adaptation.

Primary Language:Assamese (অসমীয়া) & English
Model Class:E2B Parameter Class (Base LM)
Primary Use:Pretraining Research, Fine-tuning, Perplexity/BPB Evaluation
INSTRUCTION-TUNED VARIANT
Available

AssameseLM-E2B-it

Instruction-tuned variant of AssameseLM-E2B designed for Assamese question answering, text generation, application writing, essay composition, summarisation, paraphrasing, and instruction following.

Base Model:AssameseLM-E2B
Adaptation:Supervised Instruction Tuning (Assamese SFT)
Evaluated Tasks:QA, Summarisation, Drafting, Paraphrasing, Reasoning, RAG
Planned Model Tracks (Under Active Research — Not Yet Released)
Planned
Assamese Translation

Bidirectional Assamese–English and regional Indian language neural machine translation models.

Planned
Assamese ASR

Automatic speech recognition research for conversational and dialectal Assamese audio.

Planned
Assamese TTS

Natural prosody text-to-speech synthesis for Assamese screen readers and accessibility tools.

Research / Planned
Assamese OCR

Optical character recognition and Unicode extraction for historical, educational, and administrative prints.

Scientific Focus

Principal Research Areas

AREA 01

Large Language Models

Language model adaptation, continued pretraining, supervised instruction tuning, tokenizer optimisation, and transparent benchmark evaluation.

AREA 02

Assamese NLP

Assamese text processing, morphological analysis, linguistic corpus curation, subword tokenisation, and natural language understanding.

AREA 03

Low-Resource Language Modelling

Sample-efficient pretraining, cross-lingual representation transfer, and synthetic data filtering for underrepresented regional languages.

AREA 04

Machine Translation

Assamese–English and multilingual Indian language translation systems preserving idiomatic accuracy and domain terminology.

AREA 05

Speech Technologies

Automatic Speech Recognition (ASR), Text-to-Speech (TTS), acoustic dataset curation, and voice interfaces for Assamese speakers.

AREA 06

Efficient AI Systems & OCR

Small language models, quantisation, parameter-efficient adaptation (LoRA/QLoRA), local institutional deployment, and Assamese document OCR.

Research Output

Publications & Technical Reports

All bibliographic records follow strict scientific verification policies.

TECHNICAL REPORT · MODEL DOCUMENTATION Status: Institutional Release Documentation

AssameseLM-E2B: An Open-Weight Assamese Language Model for Efficient Consumer-Grade Deployment

Authors: Dr. Pragyan Phukan, Department of Physics, Moran College, Assam, India

Abstract: While Large Language Models (LLMs) excel in high-resource languages, their performance often degrades in low-resource languages due to tokenization inefficiencies, generative instability, and linguistic interference. This paper presents AssameseLM-E2B and its instruction-tuned variant, AssameseLM-E2B-it, a 2-billion-parameter model adapted specifically for low resource Assamese language. Built upon the Gemma-4 architecture, the model incorporates an expanded byte-level BPE vocabulary to reduce tokenization overhead by nearly 500% compared with multilingual 2B--3B baselines. We employ a parameter-efficient Continual Pre-Training (CPT) pipeline with Hybrid LoRA, followed by Supervised Fine-Tuning (SFT) on a balanced, multi-domain dataset to enhance instruction-following and Retrieval-Augmented Generation (RAG) capabilities. Evaluation demonstrates a Bits-per-Byte (BPB) score of 0.7488 on FLORES-200, alongside an overall RAG score of 96.7% and 100% refusal accuracy on adversarial unanswerable prompts. Comparative analysis reveals significant limitations in general-purpose baselines, including receptive-productive decoupling, generative degradation on Romanized Assamese, and unintended Bengali script assimilation. These results demonstrate the potential of parameter-efficient adaptation for developing capable, locally deployable language models for low-resource languages under constrained computational resources.

DOI / arXiv ID: https://arxiv.org/abs/2611.06170
Academic & Ecosystem Collaboration

Let's Build the Future of Indigenous AI Together.

We invite academic researchers, universities, developers, students, linguists, Assamese language experts, and research institutions to collaborate on open datasets, model evaluation, speech technologies, and practical multilingual applications.

Joint Academic Research Student Internships & Dissertations Assamese Corpus & Dataset Contributions Local AI Deployment