This repository contains the supplementary material / appendix to go with the paper “Is Temperature the Creativity for Large Language Models” by Max Peeperkorn, Tom Kouwenhoven, Dan Brown, and Anna Jordanous.
The paper will appear at the 15th International Conference on Computational Creativity (ICCC) held in Jönköping, Sweden from 17 to 21 June 2024. It contains the code, statistical analysis, data, and generated stories, and appendix with more details regarding the survey and stories.
The appendix of the paper can be found in the supplementary_materials.pdf. This document contains the survey questions, definitions, the exemplar story and two other examples.
In the data folder, you will find the stories used in the survey in plain text format (and a metadata file), the survey results by participant and by evaluation, and all the stories generated for each temperature value (100 for each).
In the scripts folder, you will find the code we used to generated the stories and compute the embeddings.
In the analysis.ipynb notebook you will find the statistical analysis and the code that generated the figures in the paper.
Ensure that you have some version of Llama 2 Chat downloaded. This project uses llama.cpp, you will need to convert model to .guff format (and perhaps quantise if necessary, we opted for q6_k setting). The script expects the following folder structure: models/llama-2-70b-chat/ggml-model-f16.gguf.
python scripts/temperatures.py --experiment_name temperatures --model_name llama-2-70b-chat --n 100 \
--temp_min 0.001 --temp_max 2.0 --temp_n 7 --temp_scale "lin" --prompt "Write a story."When computing the embeddings, enter the model name and experiment output file you want to process, it will create a new pickle that includes the embedding vectors.
python scripts/compute_embeddings.py --experiment_path "output/temperatures.pickle" --model_name llama-2-70b-chat Snippet to be added once published.