Model post-training
A general model can understand your documents and still make the wrong decision. Post-training addresses recurring gaps in judgment, tool selection, and output quality. We work from your experts’ demonstrations and corrections to develop a specialist, then test whether that specialization holds on work the model has never seen.
Low-rank adaptation learns a weight update, ΔW = BA, while the base weights remain frozen. Domain examples define the target behavior; independent documents test whether it transfers.
The training set defines the problem the model learns to solve. We examine which decisions require domain expertise, where current models fail, and how those failures appear in real inputs. That analysis determines the example mix, the learning objective, and the evaluation split before compute is committed.
Each intervention has a purpose, a record of what changed, and a way to assess its effect.
Pair task inputs with expert solutions or ranked alternatives. Preserve the evidence available at decision time, balance common and difficult cases, and keep related documents in the same data split.
Task dataset and labeling specificationUse SFT to learn demonstrated behavior and DPO when experts can distinguish stronger responses from weaker ones. Compare LoRA adapters and full fine-tuning against the data volume, model capacity, and serving constraints.
Training configuration and checkpointCompare the specialist with the starting model using the same prompts, tools, and task budget. Examine unfamiliar inputs, rare decisions, and capabilities that must survive the update.
Held-out comparison and error analysisWe set the evaluation around the application’s requirements, including the failures an aggregate score can hide.
A task taxonomy, expert labeling rubric, dataset lineage, and a fixed baseline for the experiment.
The trained adapter or checkpoint, tokenizer, chat template, data revision, and configuration needed to reproduce the run.
The measured change in task performance, the remaining failure modes, and the next experiment or deployment recommendation.
Training teaches how to perform the task; retrieval supplies the evidence needed for a particular request. A specialist may still retrieve current documents or records. Frequently changing facts and access permissions usually belong in the retrieval layer rather than in model weights.
Enough to cover the decisions you expect the model to make, including the difficult boundaries between them. We inspect a sample, establish a baseline, and compare data subsets in an initial experiment. Those results guide further collection and labeling.
Work with the lab
Tell us where the model falls short and what better performance would mean for your team.
