Radar ·
Projection sampling for rewriting SFT training data
Paper proposes rewriting expert demonstrations as trajectories that are correct for the task and likely under the target model, before using them in standard SFT.
Radar records a paper proposing projection sampling to rewrite expert demonstrations as trajectories that are correct for the task and, at the same time, more likely under the target model. According to the source, these rewritten trajectories are then used in standard supervised fine-tuning, without changing the training method. The core idea is that adapting training examples to the model's own distribution can improve SFT. The available description does not state the paper's publication date, benchmarks, performance figures or implementation details, so these points are not asserted here.
For Rota Nacional, the topic is relevant as research context. The platform does not run this rewriting method or offer model adjustment through this path, and that capability should not be assumed. Anyone using AI to study the material should avoid pasting real expert demonstrations that contain names, CPF numbers, e-mail addresses or phone numbers. The platform detects such data and applies the organization's policy, but the best protection is to remove identifiers before sending the text. To consult the original source, search for the paper's title in academic repositories and check authors, date and results directly in the document.
Original source: Projection Sampling para Supervised Fine-Tuning