Skip to content

Latest commit

 

History

History
158 lines (120 loc) · 8.86 KB

README.md

File metadata and controls

158 lines (120 loc) · 8.86 KB
Raw

Gisting for In-Context Example Selection

This repository implements the in-context example-selection approach described in the paper GistScore: Learning Better Representations for In-Context Example Selection with Gist Bottlenecks.

Setup

  1. Clone this repository along with the icl submodule: git clone --recurse-submodules https://github.com/Shivanshu-Gupta/gist-icl/

  2. Install Python dependencies:

    pip install -r gisting/requirements.txt
pip install -r icl/requirements.txt -U

  3. Set up for ICL evaluations as described in icl/README.md.

  4. Download the finetuned and multi-task trained gist model checkpoints from here and store in gistlms/finetunes and gistlms/pretrains, respectively.

Organization

This repository is organized as follows:

gist-icl
├── gisting                 (code for training/evaluating gist models forked from https://github.com/jayelm/gisting)
├── gistlms                 (gist training logs and models)
│   ├── finetunes           (finetuned gist lms)
│   └── pretrains           (multi-task trained gist lms)
├── multitask-data          (multi-task data collections)
└── icl                     (code for ICL evaluations -- https://github.com/Shivanshu-Gupta/icl)

For details of the ICL repository, see icl/README.md.

Workflows

Finetuning Gist Models on Individual Datasets

Gist LMs are trained using gisting/src/train.py. It is a hydra script and can be run directly as python -m gisting.src.train with the parameters defined in gisting/src/arguments.py. It outputs to a directory in gistlms/ at a path configured here.

gist-train.py is a convenience wrapper of gisting/src/train.py for finetuning example gisting models on individual datasets. It defines default hyperparameters for the various datasets, constructs the command to run gisting/src/train.py and can initiate multiple runs in parallel. The process to finetune 1 and 3 token gist LMs for all the datasets used in the paper is:

  1. Output complete commands and write the parameters for all the experiments to params.jsonl:

    python gist-train.py finetune \
--lm 'flan-t5-base' \
--datasets "QNLI;MNLI;RTE;SST2;MRPC;QQP;PAWS;CMSQA;COLA;SST5;AGNEWS;SMCALFLOW_CS;MTOP;COGS;GSM8K;DROP;BOOLQ;WANLI;XNLI;MEDNLI;TWEET;PAWSX;ROTTEN_TOMATOES" \
--initckpts 'vanilla' \
--n-gists '1,3' \
--tag 'v3' --paramsfile "params.jsonl" --preview "commands"

  2. Run all the experiments in parallel on multiple GPUs:

    python icl-demo-selection/src/gist-train.py run-expsfile-parallel --paramsfile params.jsonl --gpus 0,1,2,3,4,5,6,7

The finetuned gist LMs and training logs will be stored in gistlms/finetunes at a path configured by gist-train.py:Experiment.output_dir. See gist-train.py:finetune for detailed usage. The finetuned models for all the datasets used in the paper are provided here and for individual datasets here.

To do ICL evaluation of any new gist models, update icl/src/exp_utils.py:ds2gistlms.

Multi-task Training

To train multi-task models, we first need a multi-task collection. flan_multi_task.py is used to create different subsamples of the Flan 2021, 2022 and Flan-mini collections. The subsamples are dumped in multitask-data/. The subsample used in the paper is flan2022_zs_len256_max10K which comprises up to 10K zero-shot prompts of length < 256 for each task in the Flan 2022 collection. This and many other subsamples can be downloaded from here.

The training itself is done using gisting/src/train.py directly. To train training flan-t5-large on the flan2022_zs_len256_max10K collection with 1 gist and with gradient accumulation over 64 batches of size 4 for an effective batch size of 256:

python -m gisting.src.train +model=flan-t5-large \
data.dataset_name=FLAN \
data.flan_dataset_name=flan2022_zs_len256_max10K \
training.gist.num_gist_tokens=3 \
training.gist.condition='gist' \
training.num_train_epochs=12 \
training.max_steps=-1 \
training.metric_for_best_model='eval_validation_rougeL' \
training.eval_steps=500 \
training.save_steps=500 \
data.max_eval_samples=1000 \
wandb.tag='adafactor-256-bs256' \
training.bf16=False \
training.bf16_full_eval=False \
training.lora=False \
training.per_device_train_batch_size=4 \
training.per_device_eval_batch_size=4 \
training.gradient_accumulation_steps=64 \
training.lr_scheduler_type='constant' \
training.learning_rate=5e-4 \
training.overwrite_output_dir=False \
training.optim='adafactor' training.logging_steps=50

The output directory and wandb tags for these are configured in config.yaml and gisting/src/arguments.py:global_setup(). The above will write to gistlms/adafactor-256-bs256-gist-3tok-flan-t5-large-flan2022_zs_len256_max10K. For more details see gisting/README.md or the original gisting repository.

The trained gist LMs can be directly used to gist and select in-context examples. However, to avoid dealing with the long names, the checkpoints are copied to gistlms/pretrains and then referenced in icl/src/exp_utils.py:multitask_pretrained_gistlms.

The large and xl-size gist models trained on flan2022_zs_len256_max10K with 1, 3, 6 and 15 tokens for the paper are provided here.

Running ICL Evaluations

As described above, to run ICL evaluations with the gist models, the paths to the checkpoints need to be configured in icl/src/exp_utils.py:ds2gistlms and icl/src/exp_utils.py:multitask_pretrained_gistlms. For details of how to run ICL evaluations, see icl/README.md.

Adding a new dataset

  1. Follow the steps in icl/README.md.
  2. For finetuning gist LMs some updates are necessary in gist-train.py
    1. Add it to finetune_datasets
    2. Add its TrainingParams to ds2params. Typically only bs and eval_steps need setting.
    3. Update get_metric if needed.

Command Line Tips

There are two different types of command lines in this repository:

  1. Typer - this one is used for non-nested parameterization. Allows multiple commands in a single script run as python <script> <command> <arguments>. The <command> only needs to be specified if there are more than one commands (eg. icl/src/data_params.py). The <arguments> are specified a bit differently so try running with --help to see them.
    1. gist-train.py
    2. icl/src/experiments.py:
    3. icl/src/run.py
    4. icl/src/data_params.py
  2. Hydra - this one is used for more nested parameterization.
    1. gisting/src/train.py: parameters defined in (gisting/src/arguments.py). Used to train gist LMs. Only use directly when doing multi-task training. When fine-tuning use gist-train.py.
    2. icl/src/driver.py: parameters defined in (icl/src/params.py:AllParams)

Citation

If you found this repository useful, please cite the following paper:

@article{gupta2023gistscore,
   title={GistScore: Learning Better Representations for In-Context Example Selection with Gist Bottlenecks},
   author={Shivanshu Gupta and Clemens Rosenbaum and Ethan R. Elenberg},
   year={2023},
   eprint={2311.09606},
   archivePrefix={arXiv},
   primaryClass={cs.CL}
}