Please read our Documentation: The Origin of Bunka

Bunkatopics is a package designed for Data Cleaning, Topic Modeling Visualization and Frame Analysis. Its primary goal is to assist developers in gaining insights from unstructured data, potentially facilitating data cleaning and optimizing LLMs through fine-tuning processes. Bunkatopics is constructed using well-known libraries like langchain, chroma, and transformers, enabling seamless integration into various environments.

Discover the different Use Case:

• Fine-Tuning: To achieve precise fine-tuning, it's crucial to exercise control over the data, filtering what is relevant and discarding what isn't. Bunka is a valuable tool for accomplishing this task.

• Content Overview: As an example, the Medium website offers a wealth of content across various categories such as Data Science, Technology, Programming, Poetry, Cryptocurrency, Machine Learning, Life, and more. While these categories facilitate exploration of data, they may not provide a granular overview. For instance, within the Technology category, what specific topics does Medium cover?

• Framing Analysis: Data can be analyzed in countless ways, contingent on your objectives and interests. We've developed a tool that enables you to visualize data by semantically customizing your own axes.

Theme	Google Colab Link
Visual Topic Modeling with Bunka
Cleaning dataset for fine-tuning LLM using Bunka
Understanding a dataset using Frame Analysis with Bunka
Full Introduction to Topic Modeling, Data Cleaning and Frame Analysis with Bunka.

pip install bunkatopics

git clone https://github.com/charlesdedampierre/BunkaTopics.git
cd BunkaTopics
pip install -e .

To get started, let's upload a sample of Medium Articles into Bunkatopics:

from datasets import load_dataset
docs = load_dataset("bunkalab/medium-sample-technology")["train"]["title"] # 'docs' is a list of text [text1, text2, ..., textN]

Bunkatopics offers seamless integration with Huggingface's extensive collection of embedding models. You can select from a wide range of models, but be mindful of their size. Please refer to the langchain documentation for details on available models.

from bunkatopics import Bunka
from langchain_community.embeddings import HuggingFaceEmbeddings

# Choose your embedding model
embedding_model = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2",# We recommend starting with a small model
                                        model_kwargs={"device": "cpu"}, # Or cuda if you have GPU
                                        encode_kwargs={"show_progress_bar": True}, # Show the progress of embeddings
                                        multi_process=False)  # set to True if you have mutliprocessing

# Initialize Bunka with your chosen model and language preference
bunka = Bunka(embedding_model=embedding_model, language='english') # You can choose any language you prefer

# Fit Bunka to your list of text
bunka.fit(docs)

Note: Only embedders supported by SentenceTransformers will work.

>>> bunka.get_topics(n_clusters=15, name_length=5)# Specify the number of terms to describe each topic

Topics are described by the most specific terms belonging to the cluster.

Finally, let's visualize the topics that Bunka has computed for your text data:

>>> bunka.visualize_topics(width=800, height=800, colorscale='YIGnBu')

Explore the power of Generative AI for summarizing topics! We use the 7B-instruct model of Mistral AI from the huggingface hub using the langchain framework.

from langchain.llms import HuggingFaceHub

# Define the repository ID for Mistral-7B-v0.1
repo_id = 'mistralai/Mistral-7B-v0.1'

# Using Mistral AI to Summarize the Topics
llm = HuggingFaceHub(repo_id='mistralai/Mistral-7B-v0.1', huggingfacehub_api_token="HF_TOKEN")

Note: It is recommended to use an Instruct model ie a model that has been fine-tuned on a discussion task. If not, the results might be meaningless.

# Obtain clean topic names using Generative Model
bunka.get_clean_topic_name(generative_model=llm, language='english')

Check the top documents for every topic!

>>> bunka.df_top_docs_per_topic_

Finally, let's visualize again the topics. We can chose from different colorscales.

>>> bunka.visualize_topics(width=800, height=800)

YlGnBu	Portland

delta	Blues

We can now access the newly made topics

>>> bunka.df_topics_

If you are not happy with the resulting topics, you can change them manually. Click on Apply changes when you are done. In the example, we changed the topic Cryptocurrency Impact to Cryptocurrency and Internet Discounts to Advertising.

>>> bunka.manually_clean_topics()

You have the flexibility to construct a customized dataset by excluding topics that do not align with your interests. For instance, in the provided example, we omitted topics associated with Advertising and High-Definition television, as these clusters primarily contain promotional content that we prefer not to include in our model's training data.

>>> bunka.clean_data_by_topics()

>>> bunka.df_cleaned_

The Bourdieu map provides a 2-Dimensional unsupervised scale to visualize various texts. Each region on the map represents a distinct topic, characterized by its most specific terms. Clusters are formed, and their names are succinctly summarized using Generative AI.

The significance of this visualization lies in its ability to define axes, thereby creating continuums that reveal data distribution patterns. This concept draws inspiration from the work of the renowned French sociologist Bourdieu, who employed 2-Dimensional maps to project items and gain insights.

from langchain.llms import HuggingFaceHub

# Define the HuggingFaceHub instance with the repository ID and API token
llm = HuggingFaceHub(
    repo_id='mistralai/Mistral-7B-v0.1',
    huggingfacehub_api_token="HF_TOKEN"
)

## Bourdieu Fig
bourdieu_fig = bunka.visualize_bourdieu(
        llm=llm,
        x_left_words=["This is about business"],
        x_right_words=["This is about politics"],
        y_top_words=["this is about startups"],
        y_bottom_words=["This is about governments"],
        height=800,
        width=800,
        clustering=True,
        topic_n_clusters=10,
        density=False,
        convex_hull=True,
        radius_size=0.2,
        label_size_ratio_clusters=80)

>>> bourdieu_fig.show()

positive/negative vs humans/machines	politics/business vs humans/machines

politics/business vs positive/negative	politics/business vs startups/governments

This is a beta feature. First, git clone the repository

git clone https://github.com/charlesdedampierre/BunkaTopics.git
cd BunkaTopics
pip install -e .

Then carry out a Topic Modeling and launch the serveur:

from bunkatopics import Bunka
from langchain_community.embeddings import HuggingFaceEmbeddings

# Choose your embedding model
embedding_model = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2",# We recommend starting with a small model
                                        model_kwargs={"device": "cpu"}, # Or cuda if you have GPU
                                        encode_kwargs={"show_progress_bar": True}, # Show the progress of embeddings
                                        multi_process=False)  # set to True if you have mutliprocessing

# Initialize Bunka with your chosen model and language preference
bunka = Bunka(embedding_model=embedding_model, language='english') # You can choose any language you prefer

# Fit Bunka to your text data
bunka.fit(docs)
bunka.get_topics(n_clusters=15, name_length=3)# Specify the number of terms to describe each topic

>>> bunka.start_server() # A serveur will open on your computer at https://localhost:3000/