Channel
Digests from Sebastian Raschka — Machine learning and deep learning education from Sebastian Raschka.
This deep dive into GRPO covers advanced RL techniques like clipped policy ratios and KL divergence to stabilize reasoning model training. Learn how to interpret metrics like advantage and entropy for robust model development.
Read article
This session details implementing Reinforcement Learning with Verifiable Rewards (RLVR) using the GRPO algorithm to train a reasoning model. The process involves calculating rewards, advantages, and policy gradient loss to enhance reasoning capability.
Read article
This deep dive explores inference-time scaling techniques for LLMs, covering Chain-of-Thought prompting, Temperature Scaling, Top-p filtering, and Self-Consistency. Learn how to boost model accuracy by managing compute trade-offs.
Read article
The video details building a verifier for LLM evaluation on math benchmarks. This robust pipeline involves extracting the final answer, normalizing the format, and using symbolic math libraries like sympy to verify mathematical equivalence for RL training.
Read article
This deep dive covers loading pre-trained LLMs (Qwen3), implementing autoregressive text generation in PyTorch, and optimizing performance using KV caching and torch.compile for maximum inference speed.
Read article
This guide details loading pre-trained LLMs (like Qwen3) and optimizing text generation through tokenization, KV caching, and torch.compile. Learn how to achieve high tokens/second throughput for reasoning models.
Read article
This video explains looped transformers—a technique to scale LLM depth by reusing weights multiple times. We analyze fixed repetition (Nanbeige4.2-3B) vs. dynamic Mixture-of-Recursions.
Read article
Learn how to build reasoning models from scratch, tracing the evolution from conventional LLMs to agent harnesses. The video covers setup using 'uv' and PyTorch best practices.
Read article
This deep dive explains how Anthropic's text watermarking works. It details how deterministic sampling is achieved by modifying the next-token selection process using secret keys and tournament sampling.
Read article