#llm

Short notes on attention, cross-entropy loss, and native vision transformers.

Clef and Clef-flash are Cloudflare's first self-trained decision models, hosted on Workers AI and open-weight under Apache 2.0. They return a probability for every option of typed noul, choice, and score questions, and are Jev-API compatible.

Six Apache-2.0 decision models from vLLM Semantic Router that answer choice, yes/no, and score questions in one forward pass and return a probability for every option.

Compares full fine-tuning, LoRA, and QLoRA for adapting LLMs, with pros, cons, and guidance on when to use each, plus fine-tuning tools.

Notes on choosing between prompting, RAG, and fine-tuning for LLM applications, covering PEFT, RLHF, chunking, RAG vs CAG, vector databases, and Vietnamese embedding models.

A short introduction to large language models and the three main ways to adapt them: prompt engineering, RAG, and fine-tuning.

Seven practical tips for getting useful results from ChatGPT, from giving context and using variables to splitting tasks, asking for evidence, and building conversation flows.

A taxonomy of prompting techniques with short examples, from zero-shot and few-shot prompting to chain-of-thought, multilingual, multimodal, agent-based, and answer engineering methods.

Eight prompting strategies for LLM applications, including clear instructions, reference text, task decomposition, giving the model time to think, external tools, and systematic evals, plus self-prompting.

Overview of text chunking approaches for RAG (fixed size, recursive, document-based, semantic, agentic), proposition-based retrieval, and the ChunkFormer speech model.

Explains how prompt engineering, RAG, and fine-tuning relate as ways to steer an LLM, with open-source RAG libraries and a comparison of RAG versus cache-augmented generation.

Overview of Meta's Code World Model (CWM), a 32B-parameter language model that simulates code execution, with benchmark results and download links.

Explains SENATOR, a framework that uses structural entropy, a knowledge graph, and MCTS to find LLM knowledge gaps and patch them with targeted synthetic-data fine-tuning.

NVIDIA's synthetic Vietnamese persona dataset built with NeMo Data Designer - 6M personas across 1M records (~5.4 GB), CC BY 4.0, grounded in GSO statistics.

Summary of the Viettel and NVIDIA Vietnamese LLM datasets, covering the curated corpus, the instruction set, and the function-calling benchmark.

Curated links for learning LLMs, covering prompt engineering, RAG, fine-tuning, LangChain, and model internals.

Overview of multi-agent tools and frameworks, covering Make It Heavy, Awesome LLM Apps, CrewAI, and a list of popular AI agent frameworks.

Curated list of frameworks and platforms for building multi-agent systems, including Agno, TinyTroupe, Google ADK, Dify, and TradingAgents.

Overview of vector and graph-vector databases, context stores, and RAG frameworks used for semantic retrieval in LLM applications.

Guide to JAX (autodiff, JIT, vectorization, scaling) plus a curated list of LLM application frameworks, agent tooling, and supporting libraries.

Tools for running LLMs and voice models locally, including LM Studio, Ollama, and IchigoLLaMA3.1.

Day-by-day summary of Google and Kaggle's free self-paced 5-Day Gen AI Intensive course, covering LLMs, embeddings, agents, domain-specific LLMs, and MLOps.

Learn Harness Engineering is WalkingLab's MIT course on the files, checks, and session rules around a coding agent. The English site has 14 lectures and 8 projects, and the repo was last pushed on 22 October 2026.

Notes on Tang Jie's 2026 Advanced Machine Learning course at Tsinghua, its assignments, and his statements on what comes after the chatbot era.