Skip to content

Unattended Lightweight Text Classifiers with LLM Embeddings

License

Notifications You must be signed in to change notification settings

EveripediaNetwork/fastc

Repository files navigation

Logo

Unattended Lightweight Text Classifiers with LLM Embeddings


PyPi License

Key features

  • Suitable for limited-memory CPU execution: Use efficient distilled models such as deepset/tinyroberta-6l-768d for embedding generation.
  • Logistic Regression and Nearest Centroid classification: Bypass the need for fine-tuning by utilizing LLM embeddings to efficiently categorize texts using either logistic regression or the nearest centroid through cosine similarity.
  • Efficient Parallel Execution: Run hundreds of classifiers concurrently with minimal overhead by sharing the same model for embedding generation.

Installation

pip install -U fastc

Train a model

You can train a text classifier with just a few lines of code:

from fastc import Fastc

tuples = [
    ("I just got a promotion! Feeling fantastic.", 'positive'),
    ("Today was terrible. I lost my wallet and missed the bus.", 'negative'),
    ("I had a great time with my friends at the party.", 'positive'),
    ("I'm so frustrated with the traffic jam this morning.", 'negative'),
    ("My vacation was wonderful and relaxing.", 'positive'),
    ("I didn't get any sleep last night because of the noise.", 'negative'),
    ("I'm so excited for the concert tonight!", 'positive'),
    ("I'm disappointed with the service at the restaurant.", 'negative'),
    ("The weather is beautiful and I enjoyed my walk.", 'positive'),
    ("I had a bad day. Nothing went right.", 'negative'),
    ("I'm thrilled to announce that we are expecting a baby!", 'positive'),
    ("I feel so lonely and sad today.", 'negative'),
    ("My team won the championship! We are the champions.", 'positive'),
    ("I can't stand my job anymore, it's so stressful.", 'negative'),
    ("I love spending time with my family during the holidays.", 'positive'),
    ("My computer crashed and I lost all my work.", 'negative'),
    ("I'm proud of my achievements this year.", 'positive'),
    ("I'm exhausted and overwhelmed with everything.", 'positive'),
]

Classification Kernels

Nearest Centroid

model = Fastc(
    embeddings_model='microsoft/deberta-base',
    kernel=Kernels.NEAREST_CENTROID,
)

model.load_dataset(tuples)
model.train()

Logistic Regression

from fastc import Kernels

model = Fastc(
    embeddings_model='microsoft/deberta-base',
    kernel=Kernels.LOGISTIC_REGRESSION,
    # cross_validation_splits=5,
    # cross_validation_repeats=3,
    # iterations=100,
    # parameters={...},
    # seed=1984,
)

model.load_dataset(tuples)
model.train()

Pooling Strategies

The implemented pooling strategies are:

  • MEAN (default)
  • MEAN_MASKED
  • MAX
  • MAX_MASKED
  • CLS
  • SUM
  • ATTENTION_WEIGHTED
from fastc import Pooling

model = Fastc(
    embeddings_model='microsoft/deberta-base',
    pooling=Pooling.MEAN_MASKED,
)

model.load_dataset(tuples)
model.train()

Templates and Instruct Models

You can use instruct templates with instruct models such as intfloat/multilingual-e5-large-instruct. Other models may also improve in performance by using templates, even if they were not explicitly trained with them.

from fastc import ModelTemplates, Fastc, Template

# template_text = 'Instruct: {instruction}\nQuery: {text}'
template_text = ModelTemplates.E5_INSTRUCT

model = Fastc(
    embeddings_model='intfloat/multilingual-e5-large-instruct',
    template=Template(
        template_text,
        instruction='Classify as positive or negative'
    ),
)

Save, load and export models

After training, you can save the model for future use:

model.save_model('./sentiment-classifier/')

Publish a model to HuggingFace

Important

Log in to HuggingFace first with huggingface-cli login

model.push_to_hub(
    'braindao/sentiment-classifier',
    tags=['sentiment-analysis'],
    languages=['multilingual'],
    private=False,
)

Load an existing model

You can load a pre-trained model either from a directory or from HuggingFace:

# From a directory
model = Fastc('./sentiment-classifier/')

# From HuggingFace
model = Fastc('braindao/sentiment-classifier')

Class prediction

sentences = [
    'I am feeling well.',
    'I am in pain.',
]

# Single prediction
scores = model.predict_one(sentences[0])
print(scores['label'])

# Batch predictions
scores_list = model.predict(sentences)
for scores in scores_list:
    print(scores['label'])

Inference Server

To launch the dockerized inference server, use the following script:

./server/scripts/start-docker.sh

Alternatively, on the host machine:

./server/scripts/start-server.sh

In both cases, an HTTP API will be available, listening on the fastc-server hashport 53256.

Inference

To classify text, use POST / with a JSON payload such as:

{
    "model": "braindao/tinyroberta-6l-768d-language-identifier-en-es-ko-zh-fastc-lr",
    "text": "오늘 저녁에 친구들과 함께 pizza를 먹을 거예요."
}

Response:

{
    "label": "ko",
    "scores": {
        "en": 1.0146501463135055e-08,
        "es": 6.806091549848057e-09,
        "ko": 0.9999852640487916,
        "zh": 1.471899861513275e-05
    }
}

Version

To check the fastc version, use GET /version:

Response:

{
    "version": "2.2407.0"
}