src/distilabel_dataset_generator/pipelines/textcat.py

from typing import List

import pandas as pd
from distilabel.llms import InferenceEndpointsLLM
from distilabel.steps.tasks import (
    GenerateTextClassificationData,
    TextClassification,
    TextGeneration,
)

from src.distilabel_dataset_generator.pipelines.base import (
    MODEL,
    _get_next_api_key,
)

PROMPT_CREATION_PROMPT = """You are an AI assistant specialized in generating very precise text classification tasks for dataset creation.

Your task is to write a prompt following the instruction of the user. Respond with the prompt and nothing else.

The prompt you write should follow the same style and structure as the following example prompts, clearly specifying the possible classification labels.

If a label is composed of multiple words, use a hyphen to separate them. For example, 'smartphone-review', 'customer-service', 'product-quality'.:

Classify the following customer review of a cinema as either 'positive' or 'negative'.

Classify the following news article into one or more of the following categories: 'politics', 'sports', 'technology', 'entertainment', 'health', 'business', 'environment', 'education', 'science', 'international'.

Determine the sentiment of the following social media post: 'ambiguous', 'sarcastic', 'informative', 'emotional'.

Identify the issue category for the following technical support ticket: 'billing', 'technical', 'account', 'shipping', 'returns', 'installation', 'subscription'.

Classify the following movie review into one of the following categories: 'critical', 'praise', 'disappointed', 'enthusiastic'.

Determine the level of customer satisfaction from the following customer service transcript: 'satisfied', 'dissatisfied', 'highly-satisfied', 'somewhat-dissatisfied', 'indifferent'.

Categorize the following product description into one of the following product types: 'smartphone', 'laptop', 'tablet', 'smartwatch', 'e-reader', 'headphones'.

Classify the following tweet as expressing either 'support' or 'opposition' to the political event discussed.

Classify the following restaurant review into one of the following categories: 'food-quality', 'service', 'ambiance', or 'price'.

Classify the following blog post based on its primary fashion trend or style: 'casual', 'formal', 'streetwear', 'vintage' or 'sustainable-fashion'.

User dataset description:
"""

DEFAULT_DATASET_DESCRIPTIONS = [
    "A dataset covering customer reviews for an e-commerce website.",
    "A dataset covering news articles about various topics.",
]

DEFAULT_DATASETS = [
    pd.DataFrame.from_dict(
        {
            "text": [
                "I love the product! It's amazing and I'll buy it again.",
                "The product was okay, but I wouldn't buy it again.",
            ],
            "label": ["positive", "negative"],
        }
    ),
    pd.DataFrame.from_dict(
        {
            "text": [
                "Yesterday, the US stock market had a significant increase.",
                "New research suggests that the Earth is not a perfect sphere.",
            ],
            "labels": [["economy", "politics"], ["science", "environment"]],
        }
    ),
]

DEFAULT_SYSTEM_PROMPTS = [
    "Classify the following customer review as either 'positive' or 'negative'.",
    "Classify the following news article into one of the following categories: 'politics', 'economy', 'environment', 'science', 'health'.",
]


from typing import List

MODEL = "meta-llama/Meta-Llama-3.1-8B-Instruct"


def generate_pipeline_code(
    system_prompt: str,
    difficulty: str = None,
    clarity: str = None,
    labels: List[str] = None,
    num_labels: int = 1,
    num_rows: int = 10,
) -> str:
    labels = [label.lower().strip() for label in labels or []]
    base_code = f"""
# Requirements: `pip install distilabel[hf-inference-endpoints]`
import os
from distilabel.llms import InferenceEndpointsLLM
from distilabel.pipeline import Pipeline
from distilabel.steps import LoadDataFromDicts, KeepColumns
from distilabel.steps.tasks import {"GenerateTextClassificationData" if num_labels == 1 else "GenerateTextClassificationData, TextClassification"}

MODEL = "{MODEL}"
TEXT_CLASSIFICATION_TASK = "{system_prompt}"
os.environ["HF_TOKEN"] = (
    "hf_xxx"  # https://huggingface.co/settings/tokens/new?ownUserPermissions=repo.content.read&ownUserPermissions=repo.write&globalPermissions=inference.serverless.write&canReadGatedRepos=true&tokenType=fineGrained
)

with Pipeline(name="textcat") as pipeline:

    task_generator = LoadDataFromDicts(data=[{{"task": TEXT_CLASSIFICATION_TASK}}])

    textcat_generation = GenerateTextClassificationData(
        llm=InferenceEndpointsLLM(
            model_id=MODEL,
            tokenizer_id=MODEL,
            api_key=os.environ["HF_TOKEN"],
            generation_kwargs={{
                "temperature": 0.8,
                "max_new_tokens": 2048,
            }},
        ),
        difficulty={None if difficulty == "mixed" else repr(difficulty)},
        clarity={None if clarity == "mixed" else repr(clarity)},
        num_generations={num_rows},
        output_mappings={{"input_text": "text"}},
    )
    """

    if num_labels == 1:
        return (
            base_code
            + """
    keep_columns = KeepColumns(
        columns=["text", "label"],
    )

    # Connect steps in the pipeline
    task_generator >> textcat_generation >> keep_columns

    if __name__ == "__main__":
        distiset = pipeline.run()
    """
        )

    return (
        base_code
        + f"""
    keep_columns = KeepColumns(
        columns=["text"],
    )

    textcat_labeller = TextClassification(
        llm=InferenceEndpointsLLM(
            model_id=MODEL,
            tokenizer_id=MODEL,
            api_key=os.environ["HF_TOKEN"],
            generation_kwargs={{
                "temperature": 0.8,
                "max_new_tokens": 2048,
            }},
        ),
        n={num_labels},
        available_labels={labels},
        context=TEXT_CLASSIFICATION_TASK,
        default_label="unknown"
    )

    task_generator >> textcat_generation >> keep_columns >> textcat_labeller

    if __name__ == "__main__":
        distiset = pipeline.run()
    """
    )


def get_textcat_generator(difficulty, clarity, is_sample):
    textcat_generator = GenerateTextClassificationData(
        llm=InferenceEndpointsLLM(
            model_id=MODEL,
            tokenizer_id=MODEL,
            api_key=_get_next_api_key(),
            generation_kwargs={
                "temperature": 0.8,
                "max_new_tokens": 256 if is_sample else 1024,
            },
        ),
        difficulty=None if difficulty == "mixed" else difficulty,
        clarity=None if clarity == "mixed" else clarity,
    )
    textcat_generator.load()
    return textcat_generator


def get_labeller_generator(system_prompt, labels, num_labels, is_sample):
    labels = [label.lower().strip() for label in labels]
    labeller_generator = TextClassification(
        llm=InferenceEndpointsLLM(
            model_id=MODEL,
            tokenizer_id=MODEL,
            api_key=_get_next_api_key(),
            generation_kwargs={
                "temperature": 0.8,
                "max_new_tokens": 256 if is_sample else 1024,
            },
        ),
        context=system_prompt,
        available_labels=labels,
        n=num_labels,
        default_label="unknown",
    )
    labeller_generator.load()
    return labeller_generator


def get_prompt_generator():
    prompt_generator = TextGeneration(
        llm=InferenceEndpointsLLM(
            api_key=_get_next_api_key(),
            model_id=MODEL,
            tokenizer_id=MODEL,
            generation_kwargs={
                "temperature": 0.8,
                "max_new_tokens": 2048,
                "do_sample": True,
            },
        ),
        use_system_prompt=True,
    )
    prompt_generator.load()
    return prompt_generator