Scope of work
Data preparation
Review sample provenance, usage rights, and sensitive information. Remove duplicates, conflicts, poor-quality examples, and data leakage. Separate training, validation, and held-out test sets; record dataset versions and processing rules. The project agreement defines how customer data may be used, stored, and deleted.
Supervised fine-tuning
Turn approved task inputs, target outputs, and format requirements into training examples. Select a suitable base model and training strategy. If examples are scarce, answers change frequently, or the task depends on current documents, compare retrieval and prompt changes before training.
LoRA and QLoRA
Low-Rank Adaptation trains a small set of adapter weights while keeping the base weights fixed. Quantized Low-Rank Adaptation can reduce training memory by loading the base model in a quantized form. The choice depends on model architecture, precision, hardware compatibility, and task quality; neither method is assumed to win without testing.
Adapter merging and deployment handoff
Subject to model licenses and compatibility, deliver either a separate adapter or a deployable merged model. Verify loading, inference output, and rollback. A merged model is retested; the best training-run score is not treated as a production result.
Evaluation and acceptance
Compare task accuracy, format adherence, human ratings, hallucinations, and refusal behavior under the same data split and inference settings. Check for regressions in general capability and safety behavior. The report includes dataset versions, scoring rules, failure cases, and rerun results. Acceptance thresholds are agreed before training.
Delivery is bounded by the agreed data, model, compute, and test set. We do not promise a universal improvement percentage independent of those conditions.
Start the conversation
Share the task definition, a few redacted examples, current failure cases, and available training compute. We will determine whether post-training is appropriate and propose measurable acceptance criteria.
Related services: Private LLM Deployment · Advanced Model Engineering