This repository contains all needed code to generate, preprocess, norm and evaluate semantic feature norms from GPT-3.
pip install -r requirements.txt
pip install -e .
.
├── analysis # All analyses from the paper
│ ├── category_structure # Analysis of category structure
│ ├── distribution # Analysis of featuer distribution
│ |── similarity_prediction # Analysis of human similarity judgement prediction
│ |── stats # Statistics about feature norms
│ |── dimensions # Most frequent features per dimension
│ |── quality_labling # Human judgements about sensible features
|── data
| |── gpt_3_feature_norm # Generated features
| |── decoded_answers.csv # Preprocessed and normed features -> final feature norm
| |── encoded_answers_openai.csv # Raw answers from GPT-3
| |── feature_object_matrix.csv # Feature-Concept frequency matrix based on the final feature norm
│ |── train # Priming examples for 30 runs
│ all_concepts.csv # Questions to retrieve new features
|── scripts # Different scripts
| |── data_taking # All code needed to generate features with GPT-3
| |── decoding # All code needed to preprocess and norm the feature norm
| |── vectorization # Create feature-concept matrices
To generate features with GPT-3, you need an API token from OpenAI.
cd bin
OPENAI_API_KEY python create_feature_norm.py --output_dir=. --train_dir=train --retrival_path=priming.csv
cd bin
python decode.py --answers=feature_norm_from_gpt.csv --output=. --parallel
To reproduce the analyses and figures, you can run the notebooks and scripts in the analysis directory.
Prompting - 350 Tokens Completion - 70 Tokens
GPT-4 0.03 pro Prompt 0.06 fuer completion
10 Runs - 317 317 x 10 x 350 tokens / 1000 * 0.03 = 33 317 x 10 x 70 token / 1000 * 0.06 = 13
30 Runs - 317 317 x 10 x 350 tokens / 1000 * 0.03 = 200 317 x 10 x 70 token / 1000 * 0.06 = 39
GPT-3 0.02 pro 1000 Prompt
30 Runs 317 x 420 x 30 / 1000 * 0.02 = 80
