Shivanshu-Gupta / icl-coverage

Coverage-based Example Selection

NOTE: This repository has been superseded by https://github.com/Shivanshu-Gupta/in-context-learning. Consider using that instead.

This is the repository for the paper Coverage-based Example Selection for In-Context Learning. The documentation is WIP but this should help you get started.

The code is organized as follows:

• data/ contains the datasets used in the paper. The datasets can be downloaded from here.
• src/params.py defines experiment parameters
• src/driver.py is the main file to run a single experiment. Instead of directly running this file, use src/experiments.py -- it defines default parameters and makes it easy to run multiple experiments.
• src/experiments.py contains the code to run experiments, track experiment statuses and aggregate results. Instead, of directly it dumps the parameters for all the experiments to a file that is then used by src/run.py.
• src/run.py used to run one or more experiments sequentially or in parallel on one or more gpus. It is the main file to run experiments.

Experiment results are dumped in results/.

A typical workflow is as follows:

1. This generates the parameters for 8-shot ICL with all the datasets and LLMs used in the paper and dumps them to params/params-all.jsonl.

python experiments.py --label 'final' \
--datasets "overnight;atis;smcalflow-cs;geoquery;break;mtop" --seeds '0'
--selectors "random;cosine;bm25;bertscore;set_bsr" \
--lms "cushman;codex;starcoder;neo;llama7B;llama13B" \
--lm-batch-size 20 --batch-size 20 --n-shots '8' \
--baselines-exp --paramsfile "params/params-all.jsonl" --run \
--no-collate-results

2. This runs the experiments in params/params-all.jsonl parallelly on gpus 0 and 1.

python run.py run-exps-parallel --paramsfile "params/davinci.jsonl" --gpus "0,1"