📃NLP API's

NLP API's allow to add data for analysis on LLM dashboard

Pre-requisite

• Create Project.

• Get the project_id and API key.

Initialisation

Initialise censius_client object.

Python

from censius.nlp import CensiusClient

API_KEY   = "<API KEY>"
projectId = <PROJECT ID> # datatype as integer

client = CensiusClient(api_key = API_KEY, project_id = projectId)

Introducing Types

• DatasetType

  • 🔹DatasetType.TEXT: Multi-string text containing special characters and limiters (, . & / \ * @ and so on)

• ModelType

  • 🔹ModelType.NLP

  • 🔹ModelType.LLM

• UseCase

  • NLP

    • 🔹UseCase.NLP.SUMMARIZATION

    • 🔹UseCase.NLP.SENTIMENT_CLASSIFICATION

    • 🔹UseCase.NLP.INTENT_CLASSIFICATION

    • 🔹UseCase.NLP.Q_AND_A

    • 🔹UseCase.NLP.TOXICITY_DETECTION

    • 🔹UseCase.NLP.INFORMATION_RETRIEVAL

    • 🔹UseCase.NLP.LANGUAGE_TRANSLATION

  • LLLM

    • 🔸UseCase.LLM.SUMMARIZATION

    • 🔸UseCase.LLM.SENTIMENT_CLASSIFICATION

    • 🔸UseCase.LLM.INTENT_CLASSIFICATION

    • 🔸UseCase.LLM.Q_AND_A

    • 🔸UseCase.LLM.TOXICITY_DETECTION

    • 🔸UseCase.LLM.INFORMATION_RETRIEVAL

    • 🔸UseCase.LLM.LANGUAGE_TRANSLATION

    • 🔸UseCase.LLM.REASONING

    • 🔸UseCase.LLM.VARY_PROMPTING

    • 🔸UseCase.LLM.VARY_STRATEGY

    • 🔸UseCase.LLM.CALIBRATION

    • 🔸UseCase.LLM.HARM_EVALUATION

Example Dataset

Simple dataset with News articles, reference summary, and summary generated by the T5 base model.

UseCase : Summarisation

Following APIs will be supported in the given order:

• Register Dataset : register_dataset() - Register a training dataset to the Censius Platform.

Argument	Type	Description	Presence
name	Text	A name for reference.	Required
file	CSV path	This is expected to be a Training dataset CSV file name. The CSV has to be in the provided format.	Required
dataset_type	DatasetType.TEXT	As of now by default, we are considering dataset_type as “text”. In later stage, we will be supporting: ”DatasetType.Vector”	Required
use_case	enum	UseCase.LLM.* or UseCase.NLP.* Please see the introduced types	Required

➡️ ROUGE score is calculated from generated and reference summaries. Therefore, both summaries must be provided by the user.

Python

from censius.nlp import DatasetType, ModelType, UseCase
import pandas as pd

datasetDetails = client.register_dataset(
    name="training-summarization",
    file="training-summarization.csv",
    use_case=UseCase.SUMMARIZATION,
    dataset_type=DatasetType.TEXT
)
print(datasetDetails)

datasetId = datasetDetails["dataset_id"]

Register Model - register_model()

• This API allows the user to register a new model to the Censius platform.

Argument	Type	Description	Presence
model_name	string	The name of the model	Required
model_type	enum	ModelType.NLP or ModelType.LLM; Whichever applies	Required
use_case	enum	UseCase.LLM.* or UseCase.NLP.* Introduced_types	Required
dataset_id	INTEGER	Recording the ID of the dataset the model is trained on	Required
parent_model_id	INTEGER	Id of the model being updated (version)	Optional

Python

modelDetails = client.register_model(
    model_name="summarization model",
    model_type=ModelType.LLM,
    use_case=UseCase.SUMMARIZATION, # UseCase.SENTIMENT
    dataset_id=datasetId
)


print(modelDetails)

Log - log your predictions

This function enables logging individual predictions and features. It can be integrated as part of the production environment to log below values as predictions are made.

Argument	Type	Description	Presence
log_id	string	The ID of this prediction log. This can be used to update the actual of this log later	Required
model_id	int	The ID of the model	Required
prediction	DatasetType.TEXT	The summary generated by the model	Required
referenced_output	DatasetType.TEXT	The referenced_summary used for validating prediction, hence actuals	Required
timestamp	integer	This is supposed to be Timestamp of prediction generated in millisecond.	Required
input	DatasetType.TEXT	The input query went to the LLM model.	Required
file	Pandas.DataFrame	File for bulk insertion in single call. Example provided.	Optional (WIP)
confidence_score	float	model confidence score between 0 and 1.	optional

Python

from time import time

response = client.log(
    log_id="<string>",
    model_id=<integer>,   # present in client.register_model response.
    input="<string>",
    referenced_output="<string>",
    prediction="<string>",
    confidence_score=<float> 
    timestamp=int(round(time.time() * 1000)),
)

print(response)