TRAINING-DATA INTELLIGENCE
Which data makes
your model better?
Research / Not yet available.
Our vision is to help AI teams select and evaluate datasets, examples and data mixtures based on measured downstream model performance. Training-Data Intelligence is the research pillar of our mission to make AI data traceable, trustworthy and measurably useful.
Research status
No training-data optimization product or demonstrated model improvement is available from Verifold. Our first bounded experiment did not show a benefit: the TracIn adaptation improved no distinct query and regressed one. It trained only 384 query-coordinate weights on frozen embeddings, not an encoder or LLM.
Inspect the experiment, exact results and limitations →
What we want to measure
Pre-training
Which data sources and mixtures support the intended capabilities? We want to study how coverage, repetition and filtering affect downstream performance across models and compute budgets.
Post-training
Which examples and labels help a model learn a specific task? We want to evaluate data selection for retriever fine-tuning and instruction tuning, including regressions across capability groups. We are not building our own general-purpose LLM.
Evidence before claims
Provenance, licensing, duplication and coverage matter, but they do not prove that training data improves a model. A credible comparison needs simple controls, separate final evaluation, repeated runs and the full cost of selection, training and evaluation.
Our next encoder experiment is a documented proposal and has not run. Transfer across models, customer value and commercial differentiation remain unestablished. The fourth pillar is research, not a shipping capability.