https://nalar.dev/tags/synthetic-data/ Recent content in Synthetic Data on Nalar Hugo en-us Tue, 08 Sep 2026 00:00:00 +0700 https://nalar.dev/filter-synthetic-training-data-with-rejection-sampling/ Tue, 08 Sep 2026 00:00:00 +0700 https://nalar.dev/filter-synthetic-training-data-with-rejection-sampling/ <p>Generating synthetic examples is easy; generating synthetic examples that are worth training on is harder. A language model can produce thousands of candidate answers, but blindly adding them to a training set can reinforce factual errors, weak reasoning, unwanted style, or artifacts of the generator itself.</p> <p><strong>Rejection sampling</strong> provides a simple mental model for controlling that pipeline: generate one or more candidates, evaluate each candidate with an acceptance rule, and keep only candidates that pass. The acceptance rule might use deterministic checks, a learned reward model, another language model, human review, or a combination of signals.</p>