Datasets · Fine-tuning · Model training
The data your model is missing, and the training to use it.
General-purpose models are good at general things. We build the custom and specialized datasets that make them good at yours, then fine-tune, evaluate and deploy the model. You own all of it.
What we do
From raw data to a model in production
Take one piece or the whole pipeline.
Custom dataset creation
Sourcing, cleaning, de-duplication, labeling and review, delivered in the format your training stack expects: JSONL instruction pairs, chat transcripts, preference pairs, COCO or YOLO annotations.
Specialized and domain datasets
Data a general model has never seen enough of: your industry's documents, your product's edge cases, your customers' language. Includes instruction-tuning sets, DPO/RLHF preference data, and low-resource language data such as Bangla.
LLM fine-tuning
Supervised fine-tuning and preference optimization on open-weight models (Llama, Qwen, Mistral and similar), using LoRA/QLoRA or full fine-tunes depending on budget and target quality.
Vision model training
Detection, classification and segmentation models trained on your own images. This is the same work behind our Smart QC defect-detection system, which is running in production on a manufacturing line.
Evaluation sets and benchmarks
A held-out test set built from your real use cases, so every model version gets a number instead of an opinion. We build this before training, not after.
Deployment you control
Weights and serving set up on your cloud or your own hardware, behind your access controls. No lock-in to a hosted endpoint you don't own.
How we work
Numbers early, commitment later
Every engagement starts with an evaluation set and a small pilot.
Scope
One call to pin down the task, the success metric, and whether fine-tuning is even the right tool (sometimes retrieval or better prompts win).
Data audit
We look at what you already have, what's missing, and what it will take to get it clean and labeled.
Pilot
A small dataset and a first model, measured against the evaluation set, so you see real numbers early.
Scale & deliver
Full dataset, final training runs, evaluation report, and handover of data, weights and training scripts.
We ship models in our own products too
Client work stays under NDA, but our own systems are public. Izma Smart QC runs vision models trained on a factory's own defect library, live on a manufacturing line. The Izma Office assistant uses tool-calling language models that answer only from a company's live books, under that user's permissions.
- ✓NDA before any data changes hands
- ✓You own the datasets, weights and training scripts
- ✓Labeling and training inside your environment if needed
- ✓Evaluation set agreed before training starts
- ✓Remote delivery for teams outside Bangladesh
- ✓We stay after handover, retraining as your data and needs change
Questions
Before you ask
▸Should we fine-tune a model or use retrieval (RAG)?
Use retrieval when the model needs to look up facts that change, like documents, prices or policies. Fine-tune when it needs to behave differently: follow a format, adopt a domain's vocabulary, handle a task a base model does poorly, or run smaller and cheaper. Many projects need both, and the scoping call is where we work out which.
▸Who owns the dataset and the trained model?
You do. Datasets, model weights, and training and evaluation scripts are delivered to you, and we don't reuse client data for other work.
▸Do you sign NDAs? Can you show past client work?
Yes, we sign NDAs as standard, which is also why client datasets and models aren't published on this site. We can walk you through our process and our own production systems on a call.
▸Can you work with sensitive data without it leaving our environment?
Yes. We can do the labeling and training inside your cloud account or on your hardware, with access you grant and revoke.
▸Do you work with international clients?
Yes. We're based in Dhaka, Bangladesh and work remotely with teams in other countries, with overlap hours agreed at the start of the engagement.
▸How is a project priced?
Projects are scoped per engagement, based on dataset size, labeling complexity and training compute. A small pilot first lets you judge the results before committing to the full build.
Tell us what the model gets wrong
That's usually the fastest way to scope the dataset it needs.
Start the Conversation