DecomP

Repository for Decomposed Prompting

Setup

1. Install the required dependencies

   conda create -n decomp python=3.8
   pip install -r requirements.txt

2. Set the necessary env variables

   export PYTHONPATH=src/
   export OPENAI_API_KEY=<YOUR_API_KEY>

Running Inference

Letter Concatenation

Run the letter concatenation experiments with one of the prompts

    promptid=p1 python -m decomp.inference.configurable_inference \
      --input datasets/letter_cat/n5_eg100_pos2_space.json \
      --config configs/inference/letter_cat/letter_cat_greedy_search.jsonnet \
      --output output/predictions/letter_cat_n5_eg100_pos2_space_decomp_predictions.json
      
    python -m decomp.datasets_utils.drop_eval \
      --gold_path datasets/letter_cat/n5_eg100_pos2_space.json \
      --prediction_path output/predictions/letter_cat_n5_eg100_pos2_space_decomp_predictions.json

Change promptid to p2 or p3 to try other prompts. Change config to cot_letter_cat_greedy_search to evaluate the CoT baseline and cot_rollout_letter_cat_greedy_search to evaluate the CoT w/ rollout baseline.

List Reversal

Run the list reversal experiments on sequences of length LEN using prompting strategy STRAT (STRAT should be one of decomp, cot, unrolled).

# Generate the data
mkdir -p datasets/reverse
LEN=10
STRAT=decomp
python src/decomp/datasets_utils/build_reverse_dataset.py \
  --listop=reversed \
  --elts=words \
  --out_dir=datasets/reverse
  --length=$LEN \
  --num_examples=1000

python -m decomp.inference.configurable_inference \
  --input datasets/reverse/test_${LEN}_normal_words.json \
  --config configs/inference/reverse/${STRAT}.jsonnet \
  --output output/predictions/reverse_${LEN}_${STRAT}_predictions.json
  
python -m decomp.datasets_utils.drop_eval \
  --gold_path datasets/reverse/test_${LEN}_normal_words.json \
  --prediction_path output/predictions/reverse_${LEN}_${STRAT}_predictions.json 

CommAQA

Run the CommAQA experiments with one of the prompts

    promptid=p1 python -m decomp.inference.configurable_inference \
      --input datasets/commaqa_e/iid_test_eg100.json \
      --config configs/inference/commaqa/commaqae_fine_greedy_search.jsonnet \
      --output output/predictions/commaqa_e_iid_test_eg100_fine_decomp_predictions.json
      
    python -m decomp.datasets_utils.drop_eval \
      --gold_path datasets/commaqa_e/iid_test_eg100.json \
      --prediction_path output/predictions/commaqa_e_iid_test_eg100_fine_decomp_predictions.json

Change promptid to p2 or p3 to try other prompts. Change config to cot_commaqae_greedy_search to evaluate the CoT baseline and commaqae_coarse_greedy_search to evaluate the coarse decomposition. Change the test file to datasets/commaqa_e/compgen_test_eg300.json to evaluate the compositional generalization split.

Multi-hop Open Domain QA

All the code and experiments for multi-hop open-domain QA can be found in Harsh's Decomp-ODQA repo here: https://github.com/HarshTrivedi/DecomP-ODQA