DecryptPrompt

DecryptPrompt is a Chinese-language repository that keeps updating a set of LLM resources and a series of blog posts that read papers. The resource lists cover open-source models and leaderboards, open-source frameworks for inference, fine-tuning, agents, RAG and prompting, open datasets for SFT, RLHF and pretraining, AIGC applications by field, and tutorials, blogs and conference interviews. The author also points readers to an essay titled Choose Your Weapon: Survival Strategies for Depressed AI Academics.

The core is the Decrypt Prompt series, which walks through papers topic by topic in Chinese: tuning-free prompts, freezing the prompt and fine-tuning the LM, freezing the LM and tuning the prompt, instruction tuning, automated instruction construction, LoRA fine-tuning details, RLHF comparisons across OpenAI, DeepMind and Anthropic, long inputs without training, chain of thought and many later topics. A very large paper list is organised by theme, including post-training with chain of thought and reinforcement learning, context engineering, new model architectures, mainstream LLMs and pretraining, self-evolution, RLHF, memory, multi-turn dialogue and instruction tuning. The README is about 95,000 characters. GitHub shows no licence.