Datasets & Training Recipes
A small model is only as good as the data it learns from and the recipe it is trained with. I help teams prepare and structure their datasets, blend public and private data for a specific objective, write and review training plans, and set up the version control and release cadence that keeps an internal model maintainable.
Where most training projects go wrong
Usually not in the training run. Before it.
The fine-tuning step itself is well documented and increasingly automated. What decides whether the result is useful is everything around it: whether the dataset is shaped for the task, whether the mix of sources matches what the model will see in production, whether the evaluation data is kept apart from the training data, and whether anyone can say later which data and settings produced the model now running.
This work sits alongside custom model development. I can do it as part of building a model for you, or as a review and advisory service for a team training its own.
Dataset preparation and advice
Getting the data into the right shape and the right proportions for the training objective.
Dataset preparation
Collecting, cleaning, deduplicating, labelling and formatting your data into a training-ready dataset, with a clear record of where each part came from.
Dataset structure for the task
Advice on how the dataset should be structured for the training task: classification, extraction, instruction following, structured output or speech each need a different format, balance and size.
Blending public and private data
Combining open datasets with the data you develop internally to reach a specific training objective, so the model gets breadth from public sources and specificity from yours.
Data mix review and recommendations
Reviewing an existing blend and recommending a data mix for a specific purpose: which sources to include, in what proportion, and what is missing or over-represented.
Evaluation and benchmark data
Identifying which part of the data is held out for evaluation and benchmarking, and keeping it out of the training blend so the results you report mean something.
Training recipe development and review
Drafting, updating and reviewing the plan that turns a dataset into a model.
Training plans
Drafting a training plan from scratch, or updating and reviewing one your team has written: base model, method, stages, and how each stage will be judged.
Data composition and weighting
How much each source contributes to training, whether some data should be up- or down-weighted, and how that composition changes between training stages.
The fine-tuning process
Reviewing the run itself: number of epochs, learning rate and other settings, the notebook or scripts being used, and whether the methodology being followed suits the task.
Version control, backup and release cadence
An internal model is a product your organisation maintains, not a one-off file.
Once a small model is deployed internally, it will be retrained: new data arrives, the task shifts, a better base model is released. Without version control, nobody can say which dataset and recipe produced the model in production, or roll back when a new version performs worse.
I advise on appropriate version control for models, datasets and training configurations, on where weights and data are backed up, and on a release cadence that fits how fast your task and data change, with each release checked against a fixed set of evals before it replaces the previous one.
Training a model in-house?
Send me your dataset plan or training recipe. I'll review it, or help you write one.