Way Enough - Fine-tuning gpt-3.5-turbo to learn to play "Connections"

The author attempted to fine-tune an OpenAI language model to solve the NYTimes word game "Connections". The game involves grouping 16 words into 4 categories of 4 words each. The author collected past game solutions using a script and created a dataset for fine-tuning the model. However, the initial results were not satisfactory due to errors in the dataset. After correcting these errors and re-running the fine-tuning process, the model showed slight improvement, but the results were still not impressive. The author concluded that while the experience was a valuable learning opportunity, the model did not significantly improve at the game after fine-tuning.

The author also tested the validation set with the gpt-4 model, which showed a notable improvement over the gpt-3.5 model. The author has requested access to fine-tune the gpt-4 model for future work.

Key takeaways

The author attempted to fine-tune an OpenAI language model to solve the NYTimes word game "Connections".
The initial results were not very successful, with the model struggling to correctly group words and identify categories, despite the author's efforts in data preparation and prompt engineering.
After fixing some bugs in the dataset and adjusting the model, the fine-tuned model showed a slight improvement in performance, but it was not significantly better than the original model.
The author concluded that while the results were not as expected, the process was a valuable learning experience and expressed interest in exploring fine-tuning with the gpt-4 model in the future.

Way Enough - Fine-tuning gpt-3.5-turbo to learn to play "Connections"

Key takeaways

Discussion (0)