AI Dataset Engineering
Better models begin with better datasets.
I design, generate and evaluate high-quality instruction datasets for local and domain-specific language models.
Dataset design · synthetic data · validation · fine-tuning · evaluation
GemmaQwenLlama
v22
Dataset Engineering
One shared training dataset
Gemma FT ↑Qwen FT ↑Llama FT ↑
One dataset. Multiple model families. Consistent improvements.
Research focus
The dataset is not a preprocessing step. It is the core research artifact.
Much of the current discussion around language models focuses on architecture, parameter count and compute. My work starts one step earlier: with the systematic design of the data a model learns from.
The objective is to create transparent, reproducible and transferable datasets whose effects can be measured across different model families and sizes.
The dataset is the product. Fine-tuning is the validation.
Selected evaluation results
The same dataset improved models across architectures and scales.
Average scores from the full 1,940-case evaluation. Values shown compare base and v22 fine-tuned variants.
Base
v22 fine-tuned
These figures are selected examples, not a universal claim that data always matters more than architecture. The conclusion applies to the documented training, task and evaluation conditions.
Dataset engineering pipeline
From raw knowledge to measurable model behaviour.
Source analysis
Assess structure, quality, licensing and domain constraints.
Cleaning & normalization
Transform heterogeneous material into consistent source data.
Instruction generation
Create task-specific examples with controlled output structures.
Negative examples
Add difficult cases that teach refusal, uncertainty and grounding.
Validation
Use automated checks, sampling and manual quality review.
Cross-model evaluation
Test the same dataset across architectures, scales and baselines.
Guiding principles
Reliable datasets require more than volume.
Transparency
Sources, transformations and assumptions should remain understandable.
Reproducibility
Every processing step should be documented and repeatable.
Transferability
A strong dataset should benefit more than one architecture.
Evaluation
Improvements should be tested against clear baselines and tasks.
Mission
Building better datasets for trustworthy and efficient language models.
My work combines dataset design, synthetic data generation, fine-tuning and evaluation with a focus on local AI systems, higher education and domain-specific applications.