angle-emb 0.4.0

AnglE-optimize Text Embeddings

EN | 简体中文

AnglE📐

It is Angle 📐, not Angel 👼.

📘 document: https://angle.readthedocs.io/en/latest/index.html

📢 Train/Infer Powerful Sentence Embeddings with AnglE. This library is from the paper: AnglE: Angle-optimized Text Embeddings. It allows for training state-of-the-art BERT/LLM-based sentence embeddings with just a few lines of code. AnglE is also a general sentence embedding inference framework, allowing for infering a variety of transformer-based sentence embeddings.

✨ Features

Loss:

• 📐 AnglE loss
• ⚖ Contrastive loss
• 📏 CoSENT loss
• ☕️ Espresso loss (previously known as 2DMSE, detail: README_ESE)

Backbones:

• BERT-based models (BERT, RoBERTa, ELECTRA, ALBERT, etc.)
• LLM-based models (LLaMA, Mistral, Qwen, etc.)
• Bi-directional LLM-based models (LLaMA, Mistral, Qwen, OpenELMo, etc.. refer to: https://github.com/WhereIsAI/BiLLM)

Training:

• Single-GPU training
• Multi-GPU training

More features will be added in the future.

🏆 Achievements

📅 May 16, 2024 | Paper "AnglE: Angle-optimized Text Embeddings" is accepted by ACL 2024 Main Conference.

📅 Mar 13, 2024 | Paper "BeLLM: Backward Dependency Enhanced Large Language Model for Sentence Embeddings" is accepted by NAACL 2024 Main Conference.

📅 Mar 8, 2024 | 🍞 mixedbread's embedding (mixedbread-ai/mxbai-embed-large-v1) achieves SOTA on the MTEB Leaderboard with an average score of 64.68! The model is trained using AnglE. Congrats mixedbread!

📅 Dec 4, 2023 | Our universal sentence embedding WhereIsAI/UAE-Large-V1 achieves SOTA on the MTEB Leaderboard with an average score of 64.64! The model is trained using AnglE.

📅 Dec, 2023 | AnglE achieves SOTA performance on the STS Bechmark Semantic Textual Similarity!

🤗 Official Pretrained Models

BERT-based models:

🤗 HF	Max Tokens	Pooling Strategy	Scenario
WhereIsAI/UAE-Large-V1	512	cls	English, General-purpose
WhereIsAI/UAE-Code-Large-V1	512	cls	Code Similarity

LLM-based models:

🤗 HF (lora weight)	Backbone	Max Tokens	Prompts	Pooling Strategy	Scenario
SeanLee97/angle-llama-13b-nli	NousResearch/Llama-2-13b-hf	4096	Prompts.A	last token	English, Similarity Measurement
SeanLee97/angle-llama-7b-nli-v2	NousResearch/Llama-2-7b-hf	4096	Prompts.A	last token	English, Similarity Measurement

🚀 Quick Start

⬇️ Installation

python -m pip install -U angle-emb

⌛ Load BERT-based Model

1. With Prompts: You can specify a prompt with prompt=YOUR_PROMPT in encode method. If set a prompt, the inputs should be a list of dict or a single dict with key text, where text is the placeholder in the prompt for the input text. You can use other placeholder names. We provide a set of predefined prompts in Prompts class, you can check them via Prompts.list_prompts().

from angle_emb import AnglE, Prompts
from angle_emb.utils import cosine_similarity


angle = AnglE.from_pretrained('WhereIsAI/UAE-Large-V1', pooling_strategy='cls').cuda()
# For retrieval tasks, we use `Prompts.C` as the prompt for the query when using UAE-Large-V1 (no need to specify prompt for documents).
# When specify prompt, the inputs should be a list of dict with key 'text'
qv = angle.encode({'text': 'what is the weather?'}, to_numpy=True, prompt=Prompts.C)
doc_vecs = angle.encode([
    'The weather is great!',
    'it is rainy today.',
    'i am going to bed'
], to_numpy=True)

for dv in doc_vecs:
    print(cosine_similarity(qv[0], dv))

2. Without Prompts: no need to specify a prompt. Just input a list of strings or a single string.

from angle_emb import AnglE
from angle_emb.utils import cosine_similarity


angle = AnglE.from_pretrained('WhereIsAI/UAE-Large-V1', pooling_strategy='cls').cuda()
# for non-retrieval tasks, we don't need to specify prompt when using UAE-Large-V1.
doc_vecs = angle.encode([
    'The weather is great!',
    'The weather is very good!',
    'i am going to bed'
])

for i, dv1 in enumerate(doc_vecs):
    for dv2 in doc_vecs[i+1:]:
        print(cosine_similarity(dv1, dv2))

⌛ Load LLM-based Models

If the pretrained weight is a LoRA-based model, you need to specify the backbone via model_name_or_path and specify the LoRA path via the pretrained_lora_path in from_pretrained method.

from angle_emb import AnglE, Prompts

angle = AnglE.from_pretrained('NousResearch/Llama-2-7b-hf',
                              pretrained_lora_path='SeanLee97/angle-llama-7b-nli-v2',
                              pooling_strategy='last',
                              is_llm=True,
                              torch_dtype='float16')

print('All predefined prompts:', Prompts.list_prompts())
vec = angle.encode({'text': 'hello world'}, to_numpy=True, prompt=Prompts.A)
print(vec)
vecs = angle.encode([{'text': 'hello world1'}, {'text': 'hello world2'}], to_numpy=True, prompt=Prompts.A)
print(vecs)

⌛ Load Third-party Models

You can load any transformer-based third-party models such as mixedbread-ai/mxbai-embed-large-v1, sentence-transformers/all-MiniLM-L6-v2, and BAAI/bge-large-en-v1.5 using angle_emb.

Here is an example:

from angle_emb import AnglE

model = AnglE.from_pretrained('mixedbread-ai/mxbai-embed-large-v1', pooling_strategy='cls').cuda()
vec = model.encode('hello world', to_numpy=True)
print(vec)

🕸️ Custom Train

🗂️ 1. Data Prepation

We currently support three dataset formats:

1. DatasetFormats.A: it is a pair format with three columns: text1, text2, and label (0/1).

2. DatasetFormats.B: it is a triple format with three columns: text, positive, and negative. positive and negative store the positive and negative samples of text.

3. DatasetFormats.C: it is a pair format with two columns: text, positive. positive store the positive sample of text.

You need to prepare your data into huggingface datasets.Dataset in one of the formats in terms of your supervised data.

🚂 2. Train with CLI

Use angle-trainer to train your AnglE model in cli mode.

1. Single gpu training:

Usage:

CUDA_VISIBLE_DEVICES=0 angle-trainer --help

2. Multi-gpu training:

Usage:

CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 --master_port=1234 -m angle_emb.angle_trainer --help

🚂 3. Custom Train

from datasets import load_dataset
from angle_emb import AnglE, AngleDataTokenizer


# 1. load pretrained model
angle = AnglE.from_pretrained('SeanLee97/angle-bert-base-uncased-nli-en-v1', max_length=128, pooling_strategy='cls').cuda()

# 2. load dataset
# `text1`, `text2`, and `label` are three required columns.
ds = load_dataset('mteb/stsbenchmark-sts')
ds = ds.map(lambda obj: {"text1": str(obj["sentence1"]), "text2": str(obj['sentence2']), "label": obj['score']})
ds = ds.select_columns(["text1", "text2", "label"])

# 3. transform data
train_ds = ds['train'].shuffle().map(AngleDataTokenizer(angle.tokenizer, angle.max_length), num_proc=8)
valid_ds = ds['validation'].map(AngleDataTokenizer(angle.tokenizer, angle.max_length), num_proc=8)
test_ds = ds['test'].map(AngleDataTokenizer(angle.tokenizer, angle.max_length), num_proc=8)

# 4. fit
angle.fit(
    train_ds=train_ds,
    valid_ds=valid_ds,
    output_dir='ckpts/sts-b',
    batch_size=32,
    epochs=5,
    learning_rate=2e-5,
    save_steps=100,
    eval_steps=1000,
    warmup_steps=0,
    gradient_accumulation_steps=1,
    loss_kwargs={
        'cosine_w': 1.0,
        'ibn_w': 1.0,
        'angle_w': 1.0,
        'cosine_tau': 20,
        'ibn_tau': 20,
        'angle_tau': 20
    },
    fp16=True,
    logging_steps=100
)

# 5. evaluate
corrcoef, accuracy = angle.evaluate(test_ds, device=angle.device)
print('corrcoef:', corrcoef)

💡 Others

• To enable llm training, please specify --is_llm 1 and configure appropriate LoRA hyperparameters.
• To enable billm training, please specify --apply_billm 1 and configure appropriate billm_model_class such as LLamaForCausalLM (refer to: https://github.com/WhereIsAI/BiLLM?tab=readme-ov-file#usage).
• To enable espresso sentence embeddings (ESE), please specify --apply_ese 1 and configure appropriate ESE hyperparameters via --ese_kl_temperature float and --ese_compression_size integer.
• To convert the trained AnglE models to sentence-transformers, please run python scripts/convert_to_sentence_transformers.py --help for more details.

💡 4. Fine-tuning Tips

1️⃣ If your dataset format is DatasetFormats.A, it is recommended to slightly increase the weight for cosine_w or slightly decrease the weight for ibn_w.

2️⃣ If your dataset format is DatasetFormats.B, it is recommended to set cosine_w to 0, and increase the weight for ibn_w such as 10 and 20. The angle_tau is recommended to set to 20.0.

3️⃣ If your dataset format is DatasetFormats.C, only ibn_w and ibn_tau are effective. You don't need to tune other parameters.

4️⃣ To alleviate information forgetting in fine-tuning, it is better to specify the teacher_name_or_path. If the teacher_name_or_path equals model_name_or_path, it will conduct self-distillation. It is worth to note that teacher_name_or_path has to have the same tokenizer as model_name_or_path. Or it will lead to unexpected results.

🫡 Citation

You are welcome to use our code and pre-trained models. If you use our code and pre-trained models, please support us by citing our work as follows:

@article{li2023angle,
  title={AnglE-optimized Text Embeddings},
  author={Li, Xianming and Li, Jing},
  journal={arXiv preprint arXiv:2309.12871},
  year={2023}
}

📜 ChangeLogs

📅	Description
2024 May 21	support Espresso Sentence Embeddings
2024 Feb 7	support training with only positive pairs (DatasetFormats.C)
2023 Dec 4	Release a universal English sentence embedding model: WhereIsAI/UAE-Large-V1
2023 Nov 2	Release an English pretrained model: SeanLee97/angle-llama-13b-nli
2023 Oct 28	Release two chinese pretrained models: SeanLee97/angle-roberta-wwm-base-zhnli-v1 and SeanLee97/angle-llama-7b-zhnli-v1; Add chinese README.md

📧 Contact

If you have any questions or suggestions, please feel free to contact us via email: [e-mail]

© License

This project is licensed under the MIT License. For the pretrained models, please refer to the corresponding license of the models.