Overview

This project analyzes toxicity classification and counterspeech generation using GPT-2 medium through value vector analysis.

Files and Execution Order

1. binary_classification_train.ipynb

Trains a binary classifier for toxicity detection:

• Base model: GPT-2 Medium
• Dataset: Jigsaw Toxic Comment Classification Dataset
• Output: Saves finetuned model as best_model_state.bin
• Our models can be found at Binary Classification Model.
• Metrics tracked: Accuracy, F1 Score, Loss

2. counterspeech_generation.ipynb

Generates non-toxic responses to toxic comments:

• Models used: Finetuned GPT-2 medium
• Dataset: Counter Speech Dataset for Toxic Language
• Output: Counter-speech generation model saved in saved_models directory with naming format LR_{Learning_Rate}_BS_{Batch_Size}_E_{Epochs}
• All three models can be found at Counterspeech Generation Model.

3. value_vector_analysis.ipynb

Initial value vector analysis:

• Models analyzed: Base GPT-2 Medium and fine-tuned classifier (best_model_state.bin) [Path to binary classification model can be modified if needed]
• Extracts and compares value vectors

4. analysis-1.ipynb

Deep dive into value vectors:

• Performs SVD on value vectors
• Maps vectors to vocabulary space
• Compares base vs finetuned model differences
• Uses models from steps 1 & 2

5. analysis-2.ipynb

Advanced analysis:

• Key-value vector interactions
• Cross-model vector alignments
• Token-level impact analysis
• Uses models and vectors from previous steps

Requirements

Required libraries:

pip install torch==1.9.0 transformers==4.11.3 transformer_lens==0.3.0 nltk==3.6.3 matplotlib==3.4.3 numpy==1.21.2

Models & Data

• Base Model: GPT-2 Medium (345M parameters)
• Training Data: Jigsaw Toxic Comment Dataset (~561K samples)
• Validation Data: Hold-out toxic comments (20%)
• Counter Speech Data: CONAN dataset (~15k paired examples)

Notes

• Models require GPU with 12GB+ memory
• Full training takes 9-10 hours on P100
• Save all best model checkpoints