BLOGS
Writing on language model training, evaluation, robust learning, and related statistical methods.
How continued pretraining learns from domain text and supervised fine-tuning learns from demonstrations, with an introduction to fill-in-the-middle, DeepSeek-V3's multi-token prediction, and segment-weighted losses for reasoning and final answers.
From student-generated rollouts to GOLD: how cross-tokenizer alignment enables logit distillation, with observations on early training instability and a recovery workflow using retries and segmented supervised fine-tuning.
An introduction to Group Relative Policy Optimization: how group rewards replace a learned critic, how clipping and a reference policy guide updates, and how DAPO, GSPO, and reward design extend the approach.
An intuitive introduction to Feature-wise Linear Modulation: how a conditioning input produces feature-level scale and shift controls, where FiLM fits inside neural-network blocks, and how the mechanism is trained.