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 sentence_transformers, langchain 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.
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.
# Load Embedding model
from sentence_transformers import SentenceTransformer
embedding_model = SentenceTransformer(model_name_or_path="all-MiniLM-L6-v2")
# Load Projection Model
import umap
projection_model = umap.UMAP(
n_components=2,
random_state=42)
from bunkatopics import Bunka
bunka = Bunka(embedding_model=embedding_model,
projection_model=projection_model) # the language is automatically detected, make sure the embedding model is adapted
# Fit Bunka to your text data
bunka.fit(docs)
from sklearn.cluster import KMeans
clustering_model = KMeans(n_clusters=15)
>>> bunka.get_topics(name_length=5, custom_clustering_model=clustering_model)# Specify the number of terms to describe each topic
Topics are described by the most specific terms belonging to the cluster.
topic_id | topic_name | size | percent |
---|---|---|---|
bt-12 | technology - Tech - Children - student - days | 322 | 10.73 |
bt-11 | blockchain - Cryptocurrency - sense - Cryptocurrencies - Impact | 283 | 9.43 |
bt-7 | gadgets - phone - Device - specifications - screen | 258 | 8.6 |
bt-8 | software - Kubernetes - ETL - REST - Salesforce | 258 | 8.6 |
bt-1 | hackathon - review - Recap - Predictions - Lessons | 257 | 8.57 |
bt-4 | Reality - world - cities - future - Lot | 246 | 8.2 |
bt-14 | Product - Sales - day - dream - routine | 241 | 8.03 |
bt-0 | Words - Robots - discount - NordVPN - humans | 208 | 6.93 |
bt-2 | Internet - Overview - security - Work - Development | 202 | 6.73 |
bt-13 | Course - Difference - Step - science - Point | 192 | 6.4 |
bt-6 | quantum - Cars - Way - Game - quest | 162 | 5.4 |
bt-3 | Objects - Strings - app - Programming - Functions | 119 | 3.97 |
bt-5 | supply - chain - revolution - Risk - community | 119 | 3.97 |
bt-9 | COVID - printing - Car - work - app | 89 | 2.97 |
bt-10 | Episode - HD - Secrets - TV | 44 | 1.47 |
Finally, let's visualize the topics that Bunka has computed for your text data:
>>> bunka.visualize_topics(width=800, height=800, colorscale='delta')
Explore the power of Generative AI for summarizing topics!
from langchain.llms import OpenAI
llm = OpenAI(openai_api_key = 'OPEN_AI_KEY')
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(llm=llm)
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_
topic_id | topic_name | size | percent |
---|---|---|---|
bt-1 | Cryptocurrency Impact | 345 | 12.32 |
bt-3 | Data Management Technologies | 243 | 8.68 |
bt-14 | Everyday Life | 230 | 8.21 |
bt-0 | Digital Learning Campaign | 225 | 8.04 |
bt-12 | Business Development | 223 | 7.96 |
bt-2 | Technology Devices | 212 | 7.57 |
bt-10 | Market Predictions Recap | 201 | 7.18 |
bt-4 | Comprehensive Learning Journey | 187 | 6.68 |
bt-6 | Future of Work | 185 | 6.61 |
bt-11 | Internet Discounts | 175 | 6.25 |
bt-5 | Technological Urban Water Management | 172 | 6.14 |
bt-9 | Electric Vehicle Technology | 145 | 5.18 |
bt-8 | Programming Concepts | 116 | 4.14 |
bt-13 | Quantum Technology Industries | 105 | 3.75 |
bt-7 | High Definition Television (HDTV) | 36 | 1.29 |
dataset = load_dataset("bunkalab/medium-sample-technology-tags")['train']
docs = list(dataset['title'])
ids = list(dataset['doc_id'])
tags = list(dataset['tags'])
metadata = {'tags':tags}
from bunkatopics import Bunka
bunka = Bunka()
# Fit Bunka to your text data
bunka.fit(docs=docs, ids=ids, metadata=metadata)
bunka.get_topics(n_clusters=10)
bunka.visualize_topics(color='tags', width=800, height=800) # Adjust the color
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_
doc_id | content | topic_id | topic_name |
---|---|---|---|
873ba315 | Invisibilize Data With JavaScript | bt-8 | Programming Concepts |
1243d58f | Why End-to-End Testing is Important for Your Team | bt-3 | Data Management Technologies |
45fb8166 | This Tiny Wearable Device Uses Your Body Heat... | bt-2 | Technology Devices |
a122d1d2 | Digital Policy Salon: The Next Frontier | bt-0 | Digital Learning Campaign |
1bbcfc1c | Preparing Hardware for Outdoor Creative Technology Installations | bt-5 | Technological Urban Water Management |
79580c34 | Angular Or React ? | bt-8 | Programming Concepts |
af0b08a2 | Ed-Tech Startups Are Cashing in on Parents’ Insecurities | bt-0 | Digital Learning Campaign |
2255c350 | Former Google CEO Wants to Create a Government-Funded University to Train A.I. Coders | bt-6 | Future of Work |
d2bc4b33 | Applying Action & The Importance of Ideas | bt-12 | Business Development |
5219675e | Why You Should (not?) Use Signal | bt-2 | Technology Devices |
... | ... | ... | ... |
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,
min_docs_per_cluster = 5,
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 |
---|---|
bunka.save_bunka("bunka_dump")
...
from bunkatopics import Bunka
bunka = Bunka().load_bunka("bunka_dump")
>>> bunka.get_topics(n_clusters = 15)
'''
ids = ['doc_1', 'doc_2'...., 'doc_n']
embeddings = [[0.05121125280857086,
-0.03985324501991272,
-0.05017390474677086,
-0.03173152357339859,
-0.07367539405822754,
0.0331297293305397,
-0.00685789855197072...]]
'''
pre_computed_embeddings = [{'doc_id': doc_id, 'embedding': embedding} for doc_id, embedding in zip(ids, embeddings)]
...
from bunkatopics import Bunka
bunka = Bunka()
bunka.fit(docs=docs, ids = ids, pre_computed_embeddings = pre_computed_embeddings)
from sklearn.cluster import KMeans
clustering_model = KMeans(n_clusters=15)
>>> bunka.get_topics(name_length=5,
custom_clustering_model=clustering_model)# Specify the number of terms to describe each topic
This is a beta feature. First, git clone the repository
git clone https://github.com/charlesdedampierre/BunkaTopics.git
cd BunkaTopics
pip install -e .
cd web # got the web directory
npm install # install the needed React packages
from bunkatopics import Bunka
from sentence_transformers import SentenceTransformer
embedding_model = SentenceTransformer(model_name_or_path="all-MiniLM-L6-v2")
# Initialize Bunka with your chosen model
bunka = Bunka(embedding_model=embedding_model)
# 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 http://localhost:3000/