OIHIK Lab · AI research & engineering
Advancing intelligent systems through better data, careful evaluation, and reproducible engineering. Connecting AI research with real needs in education, industry, and public service.
Machine learningLanguage AIOpen research
Reliable research starts with well-sourced corpora, documented data, and thoughtful preparation.
Research with relevance.
A shared foundation for progress.
01 / Research areas
From raw data to evaluated models, our research spans the full machine learning lifecycle. Each discipline informs the next.
Exploring how systems learn, generalize, and make useful predictions from data.
Investigating neural architectures, representations, and the methods used to optimize them.
Studying language models through training, fine-tuning, inference, and behavior analysis.
Turning diverse sources into carefully cleaned, structured, and documented research data.
Examining capabilities and limitations with repeatable benchmarks and meaningful comparisons.
Connecting data, training, evaluation, and analysis in traceable experimental workflows.
Corpora, documents, and source records
Normalization and deduplication
Quality checks and relevance
Tokenization and metadata
Pretraining and fine-tuning
Benchmarks and task evaluation
Errors, metrics, and comparisons
Refine the question and repeat
02 / Models & datasets
Our research interests connect model development with the data behind it, with a focus on Bangla, multilingual understanding, and useful evaluation.
Architectures, adaptation, and experimentation.
A large language model from OIHIK Lab.
Fine-tuning for language, tasks, and domain needs.
Representation learning and controlled comparisons.
The foundation of every meaningful experiment.
Text, grammar, examples, and language knowledge.
Bangla, English, and cross-language research data.
Curated sources, filtered data, and task benchmarks.
Release documentation will include source information, versions, intended uses, and known limitations.
03 / Evaluation
A model is more than a headline score. Our evaluation approach asks what works, where it fails, and whether another researcher can reproduce the result.
Evidence before conclusions
Method overview / evaluation criteria
| Research question | Evaluation lens |
|---|---|
| How well does it predict text? | PPL / BPB |
| Can it perform the task? | Task benchmarks |
| Is the data suitable? | Quality / coverage |
| Where does it struggle? | Error analysis |
| Can we compare fairly? | Matched settings |
04 / Research infrastructure
A connected research environment brings data, compute, models, and evaluation into the same workflow. Trace the work from its source to its findings.
05 / Projects & directions
Language resources are a starting point. Data quality and reproducible evaluation shape the research directions around them.
Collecting grammar rules, useful examples, and vocabulary in a structured workspace for language study.
Open language workspaceExploring repeatable approaches to corpus preparation, filtering, deduplication, and source documentation.
Explore the data focusDeveloping an evaluation approach that connects metrics with model behavior, context, and limitations.
Explore the approach06 / Research with public value
AI research should be understandable beyond the lab. We connect technical questions with the needs of people, institutions, and the communities they serve.
Research questions, documented methods, and reproducible experiments for students, educators, universities, and fellow researchers.
Research relevant to ministries, departments, local government, and public institutions: Bangla information access, document understanding, and evaluation of AI for public services.
Practical research directions in domain data, model adaptation, and evaluation to help engineering teams examine what works in their context.
Language resources and accessible explanations that make room for Bangla speakers, independent learners, and open research communities.
Public value starts with careful research.Data provenance, privacy, language coverage, transparent evaluation, and human oversight are questions to examine throughout the research lifecycle.
07 / Our research philosophy
Meaningful AI research takes more than training a model. It takes reliable data, repeatable experiments, careful evaluation, and the engineering to keep asking better questions.
Keep experiments repeatable, with clear configurations, data versions, and methods.
Share knowledge, tools, and findings whenever rights and responsibilities allow.
Understand the origin, coverage, and limitations of the data behind every result.
Treat infrastructure and reliable workflows as part of the research itself.
People behind the questions
A research lab is a meeting point for model researchers, data engineers, and people who understand language and real-world needs. These are the perspectives behind our approach.
Asking precise questions about how intelligent systems learn and behave.
Connecting reliable data preparation with traceable experimental workflows.
Bringing context, linguistic knowledge, and practical needs into the research.
Individual researcher profiles and publications are not yet listed.
The next question starts here
Explore the ideas, data, and engineering behind OIHIK Lab. Find the questions that connect our research with your world.
