BERT

BERT - Bidirectional Encoder Representations from Transformers
- by Google
AnswerDotAI/ModernBERT
- Apache-2.0, Python
- 2024年
- transformers v4.48.0+
- https://huggingface.co/answerdotai/ModernBERT-base
- base 22 layers, 139M 参数
- large 28 layers, 395M 参数
- ModernBERT Finally, a Replacement for BERT
  - 2024-12-19
- vs Bert
  - sequence length 512 -> 8192
  - RoPE 旋转位置嵌入支持长上下文、局部-全局交替注意力机制提高长序列处理效率、Unpadding 和 Flash Attention 技术提升推理效率
  - GeGLU 激活函数替代原始的 MLP 层以提升性能
ERNIE
urchade/GLiNER
- Apache-2.0, Python
- Generalist and Lightweight Model for Named Entity Recognition
DeBERTaV3 Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
- 2021年
google-bert/bert-base-chinese
- 2021年
https://sbert.net/
- Sentence Transformers (a.k.a. SBERT)
- text/image -> fixed-size vector representation (embedding)
- semantic textual similarity, semantic search, clustering, classification, paraphrase mining
MTEB
- https://huggingface.co/spaces/mteb/leaderboard
Cross-Encoder (a.k.a. reranker)
- re-rank top-k candidates from bi-encoder

abbr.	stand for	meaning
BERT	Bidirectional Encoder Representations from Transformers
SBERT	Sentence Transformers
MTEB	Massive Text Embeddings Benchmark
NER	Named Entity Recognition	命名实体识别
BIO	Begin, Inside, Outside	for NER
IOB	Inside, Outside, Begin	for NER
BILOU	Begin, Inside, Last, Outside, Unit	for NER
IR	Information Retrieval	信息检索
DPR	Dense Passage Retrieval	密集检索
MLDRID	in-domain (fine-tuned on the training set) evaluation
BEIR	Benchmarking Efficient Information Retrieval	高效信息检索基准
OOD	Out-of-Domain	外部领域
NLU	Natural Language Understanding	自然语言理解
GLUE	General Language Understanding Evaluation	通用语言理解评估
CSN	Code Search Net	代码搜索网络
SQA	StackQA	栈问答
BM25	Best Matching 25	最佳匹配25
MLTD	Multi-Label Text Discrimination	多标签文本判别

BIO

BIO - Begin, Inside, Outside
- 用于命名实体识别（NER）任务
- 三种标记
  - B - Begin
  - I - Inside
  - O - Outside
参考
- Inside–outside–beginning (tagging)

{
  "text": "今天天气真好",
  "entities": [
    { "start": 0, "end": 2, "label": "B-DATE" },
    { "start": 2, "end": 4, "label": "I-DATE" },
    { "start": 4, "end": 6, "label": "I-DATE" },
    { "start": 6, "end": 8, "label": "O" }
  ]
}

BIO​

BIO