This is an implementation of paper:

The Wisdom of Hindsight Makes Language Models Better Instruction Followers

If you use this code in your own work, please cite our paper:

@article{zhang2023wisdom,
  title={The Wisdom of Hindsight Makes Language Models Better Instruction Followers},
  author={Zhang, Tianjun and Liu, Fangchen and Wong, Justin and Abbeel, Pieter and Gonzalez, Joseph E},
  journal={arXiv preprint arXiv:2302.05206},
  year={2023}
}

Install BigBench

# When creating a new task, replace this with your forked repository (see below)
git clone https://github.com/google/BIG-bench.git
cd BIG-bench
python setup.py sdist
pip install -e .

Modify BIG_BENCH_DIR in utils.py to be the installation path of BigBench.

# Install other dependencies
conda env create -f conda_env.yml
conda activate hir

Modify MODEL_TYPE in utils.py to be the desired model (e.g. google/flan-t5-xl).

Change TASK to be the desired BigBench Task (e.g. logical_deduction_five_objects). Then get the results through iterative sampling and training:

bash run.sh

Please note: if running multiple experiments on one node, please assign different port numbers to different runs by changing random_port in run.sh