This directory contains necessary code to fine-tune GPT-3.5 Turbo and evaluate their safety alignment. The official APIs provided by OpenAI is used, and the only controllable hyper-parameter is the number of epochs.
-
Follow the notebooks we provided:
- tier1-harmful-examples-demonstration.ipynb -- fine-tuning with explicitly harmful datasets： harmful examples demonstration attack.
- tier2-identity-shifting-aoa.ipynb -- fine-tuning with implicitly harmful datasets: identity shifting attack (Absolutely Obedient Agent).
- tier3-benign-alpaca.ipynb -- fine-tuning with benign datasets: Alpaca
- tier3-benign-dolly.ipynb -- fine-tuning with benign datasets: Dolly
The notebooks provide examples for reimplementing our fine-tuning experiments at different risk levels, and also provide example codes for using our GPT-4 Judge to evaluate harmfulness of fine-tuned models on a few demo examples.
-
Besides, we also provide
adv_bench_evaluation.ipynbfor evaluating the safety of fine-tuned models on the public available AdvBench. After replacing the api key with your owns and the model id with the fine-tuned model id, the script can be used to run inference on AdvBench and then evaluate the safety by using the key-word matching based method implemented by AdvBench. -
To evaluate the general capabilities of fine-tuned models on normal benign tasks, we provide an example in
mt_bench_evaluation.ipynbto evaluate the performance of fine-tuned models on MT-Bench.