技法
训练、检索与适配方法，从微调、LoRA 到 RAG 与量化。每一项都指出它针对的问题，以及在算力或质量上的代价。
35 technologies
- RAG technique Retrieval-Augmented Generation: ground a model answer in documents fetched at query time.
- Embedding concept Dense numeric vector representing the meaning of a piece of text.
- Hybrid Search technique Combining keyword and vector retrieval, then merging the two rankings.
- Fine-tuning technique Continue training a pre-trained model on task-specific data to adapt its behaviour.
- RLHF technique Aligning a model with human preference using reinforcement learning on ranked outputs.
- Bi-Encoder concept Model that embeds queries and passages separately so they can be compared cheaply.
- FlashAttention technique Attention computed in tiles to avoid materialising the full matrix.
- Temperature technique Scaling the logits to make sampling more or less random.
- Supervised Fine-Tuning technique Fine-tuning a pretrained model on demonstrations of the target behaviour.
- Instruction Tuning technique Fine-tuning a model on examples of instructions and their responses.
- Chunking technique Splitting documents into passages small enough to retrieve and embed.
- Cross-Encoder concept Model that scores a query and a passage together in one forward pass.
- Grounding technique Tying a model's output to supplied sources rather than its memory.
- Top-p Sampling technique Sampling only from the smallest set of tokens whose probabilities reach p.
- Constrained Decoding technique Restricting which tokens a model may emit at each step.
- BPE technique Tokenisation that builds a vocabulary by merging frequent character pairs.
- RoPE technique Rotary position embedding, encoding position as a rotation of the query and key.
- Speculative Decoding technique Generating several tokens with a fast model and verifying them with the real one.
- QLoRA technique Fine-tuning a quantised model with LoRA adapters, so a large model fits on one GPU.
- GRPO technique Policy optimisation that scores a group of answers relative to each other.
- Reranking technique Second-stage ordering of retrieved candidates by a more accurate scoring model.
- Pretraining technique The initial large-scale training that produces a base model.
- Synthetic Data technique Training examples generated rather than collected.
- Dropout technique Randomly zeroing activations during training to reduce overfitting.