all_target_words.txt contains all the target words to play the adversarial games.

gpt4_game_top30k_results.json collects the self-play episodes of GPT-4 on the top-30K words in all_target_words.txt.

train_imitation_gpt4.json converts gpt4_game_top30k_results.json into the instruction-tuning format, with different prompt templates randomly selected from GAME_RULE_PROMPTS in utils.py.

alpaca_train.json is the SFT data from the stanford_alpaca repo. Note that you need to use the prompt templates to convert 'instruction' and 'input' to 'prompt', and 'output' to 'target'.

im_llama2_sampling_all_words_results.json includes our inference about all game episodes of the LLaMA-2 imitation-learned model.

SPAG1_sampling_all_words_results.json includes our inference about all game episodes of the LLaMA-2 SPAG-1 model.

SPAG2_sampling_all_words_results.json includes our inference about all game episodes of the LLaMA-2 SPAG-2 model.