Home

Awesome

MetaICL: Learning to Learn In Context

This includes an original implementation of "MetaICL: Learning to Learn In Context" by Sewon Min, Mike Lewis, Luke Zettlemoyer and Hannaneh Hajishirzi.

<p align="center"> <img src="img/overview.png" width="80%" height="80%"> </p> <p align="center" style="font-size: 200%"> ✨ <b><em>Check out our demo at <a href="http://qa.cs.washington.edu:2021/" target="_blank">qa.cs.washington.edu:2021</a>!</em></b> ✨ </p>

This README is mainly for how to reproduce MetaICL and Channel MetaICL in the paper, but also describe how to reproduce our baselines, including Multi-task zero-shot and various raw LM methods. All methods used in the paper are available in this repo (please see the below table).

<p align="center"> <img src="img/methods.png" width="30%" height="30%"> </p>

Updates on 02/25/2022: Our code has been updated in minor changes with (1) better preprocessing for poem_sentiment and superglue-copa datasets, and (2) some utilities added (specify --dataset for a single dataset experiment or for custom dataset).

Updates on 01/10/2022: Our code and checkpoints have been updated with a better preprocessing (using newlines instead of spaces, and removing BOS and EOS), which improves the performance by 1--4%. If you have downloaded checkpoints prior to 01/10/2022, make sure to re-download checkpoints and use the updated code. Stay tuned for the updated paper with more details and updated results. You can find the brief summary of the updated results in the Result section of this README.

For any questions about the paper or the code, please contact the first author (email) or leave issues.

If you find our code or paper useful, please cite the paper:

@inproceedings{ min2022metaicl,
  title={ Meta{ICL}: Learning to Learn In Context },
  author={ Min, Sewon and Lewis, Mike and Zettlemoyer, Luke and Hajishirzi, Hannaneh },
  booktitle={ NAACL-HLT },
  year={ 2022 }
}
<p align="center"> <img src="img/affiliations.png" width="70%" height="70%"> </p>

Content

  1. Installation
  2. Quick Start
  3. Data
  4. Training
  5. Inference
  6. Results
  7. Downloading Checkpoints

Installation

These are installation guidelines mainly for running baselines. Requirements for data are provided here. All codes are tested with Python 3.8.

pip install torch==1.9.0
pip install git+https://github.com/huggingface/transformers.git@c37573806ab3526dd805c49cbe2489ad4d68a9d7

To train the model, we use an 8-bit optimizer and mixed precision that significantly save the memory. To use them, please use the following commands (but skip if you will run inference only using released checkpoints):

# For 8-bit optimization: see https://github.com/facebookresearch/bitsandbytes for more details
pip install -i https://test.pypi.org/simple/ bitsandbytes-cuda102 # modify based on your CUDA version

# For mixed precision training: see https://github.com/NVIDIA/apex for more details
# make sure your nvcc is working (e.g. `nvcc --version`)
cd .. # move outside of this project directory
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
cd ../MetaICL # come back to this project directory

Quick Start

This is an example with a dataset financial_phrasebank.

First, prepare a list of training examples

train_data = [{"input": INPUT_1, "output": OUTPUT_1},
              {"input": INPUT_2, "output": OUTPUT_2},
              ...
              {"input": INPUT_K, "output": OUTPUT_K}]

If you prefer, you can download our training data by running the command python -m utils.download --demo_data then loading the downloaded file as follows.

with open("data/financial_phrasebank/financial_phrasebank_16_100_train.jsonl", "r") as f:
    train_data = []
    for line in f:
        train_data.append(json.loads(line))

Then, you can use our model as follows.

from metaicl.data import MetaICLData
from metaicl.model import MetaICLModel

# Load the model
data = MetaICLData(method="channel", max_length=1024, max_length_per_example=256)
model = MetaICLModel()
model.load("channel-metaicl")
model.cuda()
model.eval()

# Make a prediction for `input1`
input1 = "Both operating profit and net sales for the six-month period increased as compared to the corresponding period in 2007."
data.tensorize(train_data, [input1], options=["positive", "neutral", "negative"])
prediction = model.do_predict(data)[0]
print (prediction) # positive

# Make another prediction for `input2`
input2 = "The deal will have no significant effect on the acquiring company's equity ratio."
data.tensorize(train_data, [input2], options=["positive", "neutral", "negative"])
prediction = model.do_predict(data)[0]
print (prediction) # neutral

Data

As described in the paper, we use a collection of 142 tasks taken from CrossFit and UnifiedQA. We experiment with seven different settings, where there is no overlap in meta-training and target tasks. Download/Preprocessing guidelines are here.

Setting namealias (for command)# meta-train tasks# meta-train examples# target tasks
High Resource → Low Resourcehr_to_lr61819,20026
Classification → Classificationclass_to_class43384,02220
Non-Classification → Classificationnon_class_to_class37368,76820
QA → QAqa_to_qa37486,14322
Non-QA → QAnon_qa_to_qa33521,34222
Non-NLI → NLInon_nli_to_nli55463,5798
Non-Paraphrase Detection → Paraphrase Detectionnon_paraphrase_to_paraphrase59496,1064

To run experiments for each setting, use "alias (for command)" for commands in the Training section and the Inference section.

All settings above do not use any templates/instructions. If you want to use instruction version as in ablations in the paper, use settings in the following table.

Setting namealias (for command)# instructions / meta-train task# meta-train tasks# meta-train examples# target tasks
High Resource → Low Resource without instructionshr_to_lr_noinst032492,65512
High Resource → Low Resource with instructions (1 per task)hr_to_lr_inst132492,65512
High Resource → Low Resource with instructions (all)hr_to_lr_inst_all8.332492,65512

If you use these data resources, please make sure to cite CrossFit and UnifiedQA.

@inproceedings{ ye2021crossfit,
    title={ {C}ross{F}it: A Few-shot Learning Challenge for Cross-task Generalization in NLP },
    author={ Ye, Qinyuan and Lin, Bill Yuchen and Ren, Xiang },
    booktitle={ EMNLP },
    year={ 2021 }
}
@inproceedings{ khashabi2020unifiedqa,
    title={ {U}nified{QA}: Crossing Format Boundaries With a Single QA System },
    author={ Khashabi, Daniel and Min, Sewon and Khot, Tushar and Sabharwal, Ashish and Tafjord, Oyvind and Clark, Peter and Hajishirzi, Hannaneh },
    booktitle={ Findings of EMNLP },
    year={ 2020 }
}

If you use the instruction version, please make sure to cite the T0 paper.

@inproceedings{ sanh2022multitask,
    title={ Multitask Prompted Training Enables Zero-Shot Task Generalization },
    author={ Victor Sanh and Albert Webson and Colin Raffel and Stephen H. Bach and Lintang Sutawika and Zaid Alyafeai and Antoine Chaffin and Arnaud Stiegler and Teven Le Scao and Arun Raja and Manan Dey and M Saiful Bari and Canwen Xu and Urmish Thakker and Shanya Sharma and Eliza Szczechla and Taewoon Kim and Gunjan Chhablani and Nihal Nayak and Debajyoti Datta and Jonathan Chang and Mike Tian-Jian Jiang and Han Wang and Matteo Manica and Sheng Shen and Zheng Xin Yong and Harshit Pandey and Rachel Bawden and Thomas Wang and Trishala Neeraj and Jos Rozen and Abheesht Sharma and Andrea Santilli and Thibault Fevry and Jason Alan Fries and Ryan Teehan and Stella Biderman and Leo Gao and Tali Bers and Thomas Wolf and Alexander M. Rush },
    booktitle={ ICLR },
    year={ 2022 }
}

How to Download and Preprocess

The code is modified from the original CrossFit repo. First, install requirements:

pip install datasets==1.4.0 wget

Warning: we found that datasets==1.4.0 is not compatible with Transformers version we use for training and inference. Please use a separate environement for data preprocessing and model training/inference.

cd preprocess
# preprocess from crossfit
python _build_gym.py --build --n_proc=40 --do_test
python _build_gym.py --build --n_proc=40 --do_train # skip if you won't run training yourself
# preprocess from unifiedqa
python unifiedqa.py --do_train --do_test # skip `--do_train` if you won't run training yourself

By default, preprocessed data is saved at data/.

Additional flags:

If you want to use values that are different from default ones, please simply add the flag, e.g., python _build_gym.py --build --n_proc=40 --do_test --test_k 4.

Process instruction version

The instruction version is for settings using instructions. We use instructions from BigScience PromptSource. First, fetch instructions (prompts) from PromptSource by doing the following.

# assuming you are still inside `preprocess` directory
cd ../.. # go outside of your project directory
git clone https://github.com/bigscience-workshop/promptsource.git
cd promptsource
git checkout 4e67a38d9642bde222cb90e36e8a66fd6e4a861a
mv promptsource ../MetaICL/preprocess/ # move promptsource directory under `preprocess` directory
cd ../MetaICL/preprocess # comte back to `preprocess` directory
pip install pandas jinja2 "pyyaml>=5"

Note that this is a workaround that does not use python-pip to install the promptsource packages because it requires to use python<=3.7, while all other codes in this repo use python 3.8. If promptsource starts supporting python 3.8, please install the package following the guidelines in the original repo.

Then, download the data via:

python _build_gym.py --build --n_proc=20 --do_test --inst
python _build_gym.py --build --n_proc=20 --do_train --inst # skip if you won't run training yourself

Training

First, run the command to tensorize the text data and save them.

python train.py \
  --task $task --k 16384 --test_k 16 --seed 100 --use_demonstrations --method channel \
  --do_tensorize --n_gpu 8 --n_process 40

Then, run the following command to train the model.

python -m torch.distributed.launch --nproc_per_node=8 train.py \
  --task $task --k 16384 --test_k 16 --seed 100 --train_seed 1 --use_demonstrations --method channel --n_gpu 8 \
  --batch_size 1 --lr 1e-05 --fp16 --optimization 8bit-adam --out_dir checkpoints/channel-metaicl/$task

Training takes around 4.5 hours

If you want to train Multi-task zero-shot model that is one of our baselines in the paper, you can use similar commands for both tensorizing and training, but without --use_demonstrations and --test_k. Training takes around 3 hours.

Inference

python test.py --task $task --k 16 --split test --seed 100 --test_batch_size 16 \
    --method {channel|direct} --use_demonstrations \
    --out_dir checkpoints/metaicl/$task \
    --global_step 30000

Instead of specifying --global_step, you can specify --checkpoint for path to the checkpoint if you want to use checkpoint stored in somewhere else (for example, if you have downloaded the released checkpoints and want to use them). You must specify one of checkpoint and global_step.

If you want to run inference for Multi-task zero-shot baseline, you can use a similar command but without --use_demonstrations and --k. For this baseline, you can use --test_batch_size 64 with a 32GB GPU.

If you want to run raw LM baselines in the paper, you do not need to specify --checkpoint or --global_step. Instead, specify --do_zeroshot, and then:

You can use the same out_dir for all raw LM baselines if you are using the same GPT2 model, e.g., checkpoints/raw-gpt2-large

Results

Here is the summary of key results. Full results can be found in the paper.

Methodhr_to_lrclass_to_classnon_class_to_classqa_to_qanon_qa_to_qanon_nli_to_nlinon_paraphrase_to_paraphrase
0-shot34.834.234.240.240.225.534.2
PMI 0-shot35.133.833.840.240.227.939.2
Channel 0-shot36.537.337.338.738.733.939.5
In-context38.2/35.337.4/33.937.4/33.940.1/38.740.1/38.734.0/28.333.7/33.1
PMI In-context39.2/33.738.8/30.038.8/30.040.3/38.840.3/38.833.0/28.038.6/33.4
Channel In-context43.1/38.546.3/40.346.3/40.340.8/38.140.8/38.139.9/34.845.4/40.9
Multi-task 0-shot35.637.336.845.736.040.730.6
Channel Multi-task 0-shot38.840.942.242.136.436.835.1
MetaICL43.3/41.743.4/39.938.1/31.846.0/44.838.5/36.849.0/44.833.1/33.1
Channel MetaICL49.1/46.850.7/48.050.6/48.144.9/43.541.9/40.554.6/51.952.2/50.3

Two numbers are computed by taking the average and the worst-case performance per task (macro-F1 for classification tasks and accuracy for others) over five different seeds, and then taking the macro-average over tasks (one number for 0-shot models since they are not dependent to the seed). Bold indicates the best average result.

Updates on 01/10/2022: The results are updated from the previous version with a better preprocessing. Updates are not reflected in the paper yet---stay tuned!

Downloading Checkpoints

You can run the inference script by specifying --checkpoint {model_name}, and the script will automatically download the corresponding checkpoint under the checkpoints/ directory. {model_name} can either be

Alternatively, you can download all checkpoints via:

python -m utils.download --checkpoints --setting all --method all

If you want to download one of settings only, specify --setting {setting_name} (using "alias for command" in the setting table above) If you want to download one of methods only, specify --method {method_name} where method_name is one of metaicl, channel-metaicl, multitask-zero, channel-multitask-zero.

Simply reproducing all results in the paper

You can use the following commands (based on a 32GB GPU):

# raw LM zero-shot baselines (0-shot, PMI 0-shot, Channel 0-shot)
bash reproduce.sh {setting_name} {zero|pmi-zero|channel-zero} 100 64

# raw LM in-context baselines (in-context, PMI in-context, Channel in-context)
bash reproduce.sh {setting_name} {ic|pmi-ic|channel-ic} 100,13,21,42,87 16

# Multi-task 0-shot baselines
bash reproduce.sh {setting_name} {multitask-zero|channel-multitask-zero} 100 64

# MetaICL
bash reproduce.sh {setting_name} {metaicl|channel-metaicl} 100,13,21,42,87 16

License

MetaICL is CC-BY-NC 4.0 licensed.