LLM
大语言模型：一种在文本上训练的 Transformer，其规模足以产生通用能力。
概览
大型语言模型是一种 Transformer 网络，在超大规模文本语料上训练以预测下一个 token。它解决了构建一个无需任务特定训练数据即可执行多种语言任务的系统的问题。它之所以存在，是因为语言任务上的能力被证明会随模型规模、数据和算力可预测地提升，这使得一个通用文本模型比一批狭窄专用模型更值得投入。本图谱中的大多数 AI 技术，要么是为了让 LLM 适配某个领域，要么是为了降低其运行成本。
技术图谱
LLM 如何与技术栈的其他部分相连。
使用与依赖
- Tokenization technique Techniques LLM uses Tokenization High ✓ source
- Chatbot application Applications Chatbot based on LLM High ✓ source
- Claude model Models Claude based on LLM Medium ✓ source
- Code Generation application Applications Code Generation based on LLM High ✓ source
- Few-Shot Learning technique Prompting Few-Shot Learning based on LLM High ✓ source
- Gemini model Models Gemini based on LLM Medium ✓ source
- In-Context Learning concept Prompting In-Context Learning based on LLM High ✓ source
- LLaVA model Multimodal LLaVA based on LLM High ✓ source
- Multi-Agent technique AI Agents Multi-Agent based on LLM High ✓ source
- Query Rewriting technique Techniques Query Rewriting based on LLM High ✓ source
- Transformer architecture Algorithms LLM based on Transformer High ✓ source
- Zero-Shot concept Prompting Zero-Shot based on LLM High ✓ source
被用于
- AI Agent application AI Agents LLM used by AI Agent High ✓ source
- GraphRAG technique Techniques LLM used by GraphRAG High ✓ source
- RAG technique Techniques LLM used by RAG High ✓ source
应用于
- Alignment technique Safety & Alignment Alignment used for LLM High ✓ source
- B200 hardware Hardware B200 used for LLM High ✓ source
- Beam Search technique Techniques Beam Search used for LLM High ✓ source
- Distillation technique Techniques Distillation used for LLM Medium ✓ source
- GPQA benchmark Benchmarks GPQA used for LLM High ✓ source
- GSM8K benchmark Benchmarks GSM8K used for LLM High ✓ source
- H100 hardware Hardware H100 used for LLM High ✓ source
- H200 hardware Hardware H200 used for LLM High ✓ source
- HellaSwag benchmark Benchmarks HellaSwag used for LLM High ✓ source
- HumanEval benchmark Benchmarks HumanEval used for LLM High ✓ source
- KV Cache technique Infrastructure KV Cache used for LLM High ✓ source
- LMArena benchmark Benchmarks LMArena used for LLM High ✓ source
- MMLU benchmark Benchmarks MMLU used for LLM High ✓ source
- MT-Bench benchmark Benchmarks MT-Bench used for LLM High ✓ source
- MoE architecture Algorithms MoE used for LLM High ✓ source
- Pretraining technique Techniques Pretraining used for LLM High ✓ source
- Prompt Engineering technique Prompting Prompt Engineering used for LLM High ✓ source
- Quantization technique Techniques Quantization used for LLM High ✓ source
- SQuAD dataset Datasets SQuAD used for LLM High ✓ source
- Softmax concept Algorithms Softmax used for LLM High ✓ source
- TPU hardware Hardware TPU used for LLM High ✓ source
- Temperature technique Techniques Temperature used for LLM High ✓ source
- TensorRT-LLM tool Infrastructure TensorRT-LLM used for LLM Medium ✓ source
- The Pile dataset Datasets The Pile used for LLM High ✓ source
- Tokenizer tool Infrastructure Tokenizer used for LLM High ✓ source
- Top-p Sampling technique Techniques Top-p Sampling used for LLM High ✓ source
- vLLM tool Infrastructure vLLM used for LLM High ✓ source
- Transformer architecture Algorithms Transformer enables LLM High ✓ source
Classification
- Context Window concept Models Context Window part of LLM Medium ✓ source
- Hallucination concept Models Hallucination part of LLM High ✓ source
- Inference concept Infrastructure Inference part of LLM High ✓ source
- Logits concept Models Logits part of LLM High ✓ source
- Token concept Models Token part of LLM High ✓ source
- Tokenizer tool Infrastructure Tokenizer part of LLM High ✓ source