Borrow from OpenAI/weak-to-strong

[abdurraheemali]

https://github.com/openai/weak-to-strong changes a lot of tasks into binary prediction, we could take a subset of those where the prediction can influence the outcome, load small pretrained models, do full finetuning with the zero sum training method, and measure what effect it has.

[abdurraheemali]

some materials also in https://github.com/johannestreutlein/scoring-rules-performative

[abdurraheemali]

https://github.com/jcperdomo/performative-prediction/blob/main/experiments/neurips2020/stochastic-credit-simulator.ipynb

[abdurraheemali]

the training script they have is

python train_weak_to_strong.py --batch_size 32 --max_ctx 512 --ds_name "sciq" --loss "logconf" --n_docs 1000 --n_test_docs 100 --weak_model_size "gpt2-medium" --strong_model_size "gpt2-large" --seed 42

our changes:

• weak model will be the same as strong model
• we will estimate whatever can be trained on 1xH100 on short notice and just do those (I have runpod credits but might be easier to use modal labs)
• gotta filter to a subset of the datasets (not clear on this yet)