Skip to content

[NeurIPS 2022] Generating Training Data with Language Models: Towards Zero-Shot Language Understanding

License

Notifications You must be signed in to change notification settings

yumeng5/SuperGen

Repository files navigation

SuperGen

The source code for Generating Training Data with Language Models: Towards Zero-Shot Language Understanding, published in NeurIPS 2022.

Requirements

Before running, you need to first install the required packages by typing following commands (Using a virtual environment is recommended):

pip3 install -r requirements.txt

Overview

SuperGen is a Supervision Generation method for zero-shot learning on NLU tasks. Instead of training on task-specific data, SuperGen generates training data guided by label-descriptive prompts with a unidirectional language model and fine-tunes another language model on the generated data.

Training and Test Data: Our method does not use any task-specific data (e.g., original training set). We provide our generated training set and original dev set (used as the test set) of each GLUE task under the data directory: train.json files are the generated training set (after data selection); test.tsv files are the original GLUE dev set (used as the test set for evaluation purpose).
Pretraining Corpus: We provide the processed pretraining corpus (Wikipedia and OpenWebText) for generating training data for sequence-pair tasks under the pretrain_corpus directory; see the README file there for details.

Generating Training Data

The generated training set used in the paper are provided as train.json files under each task directory; you should be able to obtain very similar generated data by following the steps below:

Data Generation: The entry script for generating training data for GLUE tasks is gen_train_data.py. The basic usage is

python gen_train_data.py --task $TASK --label $LABEL --save_dir $SAVE_DIR --num_gen $NUM_GEN

You can generate training data of each label either by setting individual label name $LABEL one at a time or by setting $LABEL=all to generate data for all labels (this will still be done sequentially). You may want to set $NUM_GEN to be larger than the desired training set size, as only those texts with the highest generated probability will be used to form the final training set.

Data Selection: After generating the training data, the final training set can be constructed by running the following:

python src/gen_utils.py --task $TASK --num_select_samples $NUM_SELECT \
                        --read_dir $SAVE_DIR --save_dir $DATA_DIR

Example: We provide an example script run_gen.sh that includes the entire generation process for all GLUE tasks under the setting described in the paper.

Fine-Tuning

The entry script for fine-tuning on generated data is finetune.py. The basic usage is

python finetune.py \
    --task_name $TASK \
    --data_dir data/$TASK \
    --overwrite_output_dir \
    --do_train \
    --do_predict \
    --smooth $SM \
    --momentum $MOMENT \
    --eval_steps $INTERVAL \
    --threshold $TH \
    --reg_weight $REG \
    --temp_ensemble_rampup $RAMP \
    --model_name_or_path $MODEL \
    --max_seq_length 128 \
    --first_sent_limit 100 \
    --per_device_train_batch_size $BS \
    --learning_rate $LR \
    --num_train_epochs 3 \
    --output_dir $OUT_DIR \
    --template $TEMPLATE \
    --mapping $MAPPING \
    --warmup_ratio 0.1 \
    --save_at_last \

Example: We provide an example script run_finetune.sh with command line arguments set up for all GLUE tasks under the setting described in the paper.

Results: When using the same prompt-based fine-tuning pipeline (with the same manual prompts and label words), zero-shot SuperGen even achieves better performance than few-shot LM-BFF using 32 annotated samples per class across seven GLUE classification tasks:

Method MNLI-m/mm QQP QNLI SST-2 CoLA RTE MRPC AVG
LM-BFF 32-Sample Few-Shot 68.3/70.5 65.5 64.5 92.7 9.3 69.1 74.5 63.6
SuperGen Zero-Shot 72.3/73.8 66.1 73.3 92.8 32.7 65.3 82.2 69.4

Acknowledgement

Some scripts in this repository are adapted from COCO-LM (for COCO-LM model), LM-BFF (for prompt-based fine-tuning) and huggingface transformers (for text generation and GLUE processor/trainer).

Citations

Please cite the following paper if you find the code helpful for your research.

@inproceedings{meng2022generating,
  title={Generating Training Data with Language Models: Towards Zero-Shot Language Understanding},
  author={Meng, Yu and Huang, Jiaxin and Zhang, Yu and Han, Jiawei},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}

Releases

No releases published

Packages

No packages published