Anthropic Researcher explains how to train LLMs

Q2Qwen 2.5 0.5BAlibabaAvailable in EU

Used in a free Google Colab demo to show the difference between a pre-trained base model (pure autocomplete) and an instruct model trained on assistant-formatted data with a stop token.

OOOpenAI o1OpenAIAvailable in EU

Described as the first public thinking model, trained with reinforcement learning to reason through trial and error. It improved on tasks like ARC and coding and hallucinated less.

DRDeepSeek R1DeepSeekAvailable in EU

Its paper showed publicly how to train reasoning with reinforcement learning using GRPO (group relative policy optimization). GRPO samples many answers per problem, scores them against the group average (advantage), and reinforces better-than-average answers.