LLM Course

This repository contains tutorials and scripts for working with Large Language Models (LLMs), natural language processing, and text analysis.

Tutorials (Jupyter Notebooks)

🔒 deidentification.py

Text De-identification using Presidio

• Remove personally identifiable information (PII) from text
• Uses Microsoft Presidio Analyzer and Anonymizer
• Detects and anonymizes names, phone numbers, and other sensitive data
• Leverages spaCy's NLP models for entity recognition
• Batch processing support with progress tracking using tqdm

📊 openrouter_api.ipynb

Measure constructs using LLMs

• Learn how to use the OpenRouter API to access various LLM models
• Process API outputs and extract structured data from LLM responses
• Measure psychological and social constructs using LLMs
• Save results to CSV for further analysis
• Includes error handling and batch processing examples

GUI version (in progress): https://huggingface.co/spaces/freymon-perez/construct_encoding_demo

🔗 embeddings_tutorial.ipynb

• Generate text embeddings using Spanish transformer models (sentence-transformers)
• Apply UMAP for dimensionality reduction to visualize in 2D
• Perform clustering on vector representations
• Visualize semantic relationships with interactive plots using Plotly
• Demonstrates how embeddings capture semantic meaning across different social constructs

📝 linguistic_properties.ipynb

English Linguistic Features with spaCy + TextDescriptives

• Extract comprehensive linguistic and readability features from English text
• Use spaCy and TextDescriptives library for text analysis
• Integrate TextDescriptives as a spaCy pipeline component
• Process multiple texts and export metrics to DataFrame
• Includes tokenization, POS tagging, dependency parsing, and readability scores

📝 propiedades_linguisticas_spacy_es.ipynb

Propiedades Lingüísticas en Español

• Extract linguistic features from Spanish text using spaCy
• Analyze tokens, lemmas, part-of-speech tags, and morphology
• Parse syntactic dependencies and named entities
• Traditional NLP approach without requiring API keys
• Demonstrates when to use spaCy vs. LLMs for linguistic analysis

🎭 sentiment_analysis_emotion_recognition.ipynb

Sentiment Analysis and Emotion Classification (Pre-trained)

• Use pre-trained transformer models for sentiment analysis
• Classify emotions in text using RoBERTa-based models
• Leverage Hugging Face transformers library and pipelines
• GPU acceleration support for faster processing
• Apply models to analyze emotional content in text data

🌐 reddit_download.ipynb

Reddit Data Collection

• Download posts and comments from Reddit using the PRAW library
• Search and filter subreddits based on topics
• Extract metadata including timestamps, scores, and author information
• Save Reddit data to structured formats for analysis
• Includes rate limiting and error handling

Discover constructs

GUI version (in progress): https://huggingface.co/spaces/freymon-perez/groupNLPExplanations

🔑 api_keys.py

API Key Configuration

• Stores API keys for various services (OpenRouter, Reddit, etc.)
• Note: Keep this file private and never commit it to public repositories
• Add to .gitignore to prevent accidental exposure

Data

data/input/

Codebooks

• suicide_risk_lexicon_codebook_prototypical_examples.txt from the Suicide Risk Lexicon
• meditation_codebook.txt Created one with GPT-5

Datasets

• reddit_df_subset_random_2500.csv 2500 posts from Mindfulness, ADHD, Lonely, and Anxiety subreddits, a subset of this dataset https://mair.sites.fas.harvard.edu/datasets/rmhd_27subreddits_1300posts_train.csv
• meditation_interview_gpt5.txt generated an interview with GPT-5 on the meditation experience of a person
• data/input/datasets/goemotions_200-447.csv a subset of Google's GoEmotions dataset (single sentences from Reddit labelled by humans for 28 emotions) with this prompt:

  Create a codebook like this one but for about 20-30 constructs relevant to mindfulness (positive and negative affect, 5 hindrances, insight, some mental health symptoms, barriers, etc (these include definitions and examples of things people might say in first person), list one under the other without bullet points or blank lines, one line per construct:
  
  

Impulsive: A predisposition toward rapid, unplanned reactions to internal or external stimuli without regard to the negative consequences of these reactions to the impulsive individuals or to others. Examples: act on impulses; act without thinking; impulsive; impulsivity; rash decision; reckless; say things without thinking; snap decision; spur of the moment; urge Defeated and feeling like a failure: The perception of a failed struggle, feelings of powerlessness and a sense of losing social status or missing personal goals. Examples: I feel like a failure; I'm a failure; I'm a loser; am powerless; crushed; defeated; demoralized; feel powerless; humiliated; inferior ```

data/output/

• openrouter_results.csv - Results from LLM API calls on Reddit data

Getting Started running on your local computer

1. Open terminal, create and activate the virtual environment:

python -m venv llm_course
source llm_course/bin/activate  # On macOS/Linux
# or
llm_course\Scripts\activate  # On Windows

2. Install required dependencies:

pip install sentence-transformers umap-learn plotly pandas
pip install spacy textdescriptives
pip install transformers torch
pip install praw
pip install presidio-analyzer presidio-anonymizer
python -m spacy download en_core_web_lg
python -m spacy download es_core_news_sm

3. Configure API keys in api_keys.py (if needed)

4. Open Jupyter notebooks