search

forrestdavis / NLPScholar

Tools for training an NLP Scholar
GNU General Public License v3.0
5 stars 2 forks source link

readme

NLPScholar

This repository is used for NLP at Colgate University, taught by Profs. Forrest Davis and Grusha Prasad. It is mainly a toolkit for running various NLP experiments with the aim of training NLP Scholars! It builds on top of the wonderful HuggingFace NLP tools.

Never heard of NLP Scholars before, no worries, we wrote a paper: Training an NLP Scholar at a Small Liberal Arts College: A Backwards Designed Course Proposal which we presented at the Sixth Workshop on Teaching NLP co-located with ACL 2024.

PLEASE NOTE: This is under active development. If you run into issues, we are sorry. First thank you so much for using this. We would love to learn that you are and how you incorporate it in your class/research. You can cite this in your work if you want (see the bottom) :). Second, please submit an issue on github. We will try to promptly fix it (if its not a crazy time of the semester). If you are in our class, please email us.

Now onto the details!

Environment

Install the nlp environment (see Install.md). Ensure you have run 

conda activate nlp

Running

You can run experiments via main.py with the relevant config file. For example, you can try a sample config for interact:

python main.py sample_configs/interact.yaml

If no config is provided as an argument, config.yaml is used. See below for details on the structure of the config files.

Basic Config Files

Interacting with the toolkit is facilitated by config files. An example one is copied below showing how to run GPT2 in interactive mode. 

exp: MinimalPair

mode: 
    - interact

models: 
      hf_causal_model:
          - gpt2

exp

There are three experiments: MinimalPair, TextClassification, and TokenClassification. MinimalPair is for by-token (that is, including subword) predictability measures for minimal pair experiments (i.e., targeted syntactic evaluations). TextClassification is for classification over texts. In other words, one label is returned for each text inputted (e.g., sentiment analysis, natural language inference). TokenClassification is for classification over tokens in a text. In other words, one label is returned for each token (i.e., subword) in a text (e.g., part of speech tagging, named entity recognition).

mode

Each experiment has four modes: interact, train, evaluate, and analyze.

Experiment Interact Train Evaluate Analyze
MinimalPair Returns by-word predictability for an inputted sentence (combining subword tokens) Finetunes LM using relevant objective (e.g., autoregressive, MLM) Returns by-token predictability measures for each input in a dataset Returns average difference in predictability between expected and unexpected for each condition and model
TextClassification Returns by-text classification labels for some inputted text Finetunes a pretrained model for text classification Returns by-text labels for each input in a dataset Returns average accuracy, precision and recall for each class in each condition
TokenClassification Returns by-token classification labels for an inputted text Finetunes a pretrained model for token classification Returns by-token labels for each input in a dataset Returns average accuracy, precision and recall for each class in each condition

models

There are 4 types of models supported (all building on HuggingFace's transformers library): causal language models (hf_causal_model), masked language models (hf_masked_model), text classification models (hf_text_classification_model), and token classification models (hf_token_classification_model). You can use any model that is listed on HuggingFace's models hub as long as they work with the relevant auto classes (AutoModelForCausalLM, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification). To refer to a model, you give it's type and then it's model name on HuggingFace (or path to the folder on your local computer). More than one model from each class can be specified. For example, 

models: 
      hf_causal_model:
          - gpt2
      hf_masked_model:
          - bert-base-cased
          - roberta-base

Will load one causal language model (GPT2) and two masked language models (BERT and RoBERTa).

Config Details for interact

The interact mode builds on the evaluate mode and allows you to interact with a model before running it. You simply specify a model (and any additional config settings): 

exp: MinimalPair

mode: 
    - interact

models: 
      hf_masked_model:
          - bert-base-cased

Config Details for evaluate

datafpath and predfpath

In running in evaluate mode, you need to specify the path to the file with the data you will be running the model on (the datafpath) and the path to where you want to save the predictions (the predfpath). For example, 

exp: MinimalPair

mode:
    - evaluate

models: 
      hf_causal_model:
          - gpt2
      hf_masked_model:
          - bert-base-cased

datafpath: data/minimal_pairs.tsv

predfpath: predictions/minimal_pairs.tsv

Runs GPT2 and BERT on the minimal pairs in data/minimal_pairs and saves the output to predictions/minimal_pairs.tsv.

checkFileColumns

In evaluating models (in evaluate mode), you can check that the necessary columns for the broader experiment are included in the file indexed with datafpath. The default value is True, meaning the columns are checked. 

checkFileColumns: True

loadAll

In evaluating models (in evaluate mode), more than one model can be evaluated. loadAll controls memory usage, with True for loading all models to memory at once and False for loading one model at a time into memory. The default behavior is False. 

loadAll: False

stride

When using models with a fixed context length (like gpt2), care needs to be taken with calculating the predictability measures of a token in long contexts which exceed the maximum length of the model. In these cases, we use a striding window strategy. You can use stride to control the size of the stride. The default behavior is half of the maximum length allowed by the model. 

stride: 100

To make this clear, here is an example. Suppose the maximum context length was 5, each word was represented as one token, and our stride was 3. The following sentence would cause problems:

the strange boy is outside and the girl saw him

As it has 10 words, but the model only allows for 5. Using a stride would yield the following fragments, over which predictability measures are calculated:

the strange boy is outside

is outside and the girl

the girl saw him

The first occurrence of the word is where it's probability is calculated. So for example, outside is P(outside | the strange boy is) and girl is P(girl | is outside and the).

Note that for certain models, special tokens are appended to the start and end of a text (like [CLS]). At the moment, we handle this just for masked language models. Importantly, these additional tokens interact with stride, as we have implemented it. They are ignored in calculating the stride jumps and maximum context length. So, in effect, the maximum length is less than the total allowed (by two for now, though future versions should handle this more rigorously).

Config Details for analyze

The analyze mode is designed to take the predictions from the evaluate mode and generate summaries that are relevant to each experiment type. More details can be found in the following markdown files in src/analysis/: MinimalPairAnalysis.md, TextClassificationAnalysis.md, and TokenClassificationAnalysis.md.

In the analyze mode, you need to specify three filepaths:

Each experiment type has several other optional parameters, the details of which can be found in the experiment specific markdown files in src/analysis/.

Config Details for train

trainfpath, validfpath, and modelfpath

In train mode you need to specify training data, validation data, and an output directory for the final model. Training and validation data can either be from HuggingFace's dataset options on their hub or local json or tsv files. To specify a HuggingFace remote dataset specify the name, task (if applicable), and split with colon seperators. See below for two examples, one loading imdb data, which doesn't have subtasks: 

trainfpath: imdb:train
validfpath: imdb:test
modelfpath: imdb_model

and one loading mnli from glue: 

trainfpath: nyu-mll/glue:mnli:train
validfpath: nyu-mll/glue:mnli:validation_matched
modelfpath: mnli_model

loadPretrained

You can specify if you want to load the pretrained weights of a model or randomly initialize a model with the same architecture as the model name with loadPretrained. If set to False and training for classification you must specify the number of labels with numLabels. The default is True. 

loadPretrained: True

numLabels

You can specify the number of classification labels for token or text classification with numLabels. Note: you must specify a value if loadPretrained is False. 

numLabels: 5

maxTrainSequenceLength

In loading a model, you can specify the maximum sequence length (i.e., the context size) with maxTrainSequenceLength. This changes the sequence length of the model when loading not from a pretrained model and controls the sequence length in a batch during training. The default is 128. 

maxTrainSequenceLength: 128

seed

You can specify the seed for shuffling the dataset initially with seed. The default value is 23.

seed: 23

samplePercent

It is often helpful to run training code with a subset of your data (e.g., for debugging). You can specify what percent of your data to use with samplePercent. This can either be a whole number between 0 and 100 (which is converted to a percent; for example, 10 translates to 0.10), or a float (e.g., 0.001). The default is None, which results in no sampling (i.e., all data is used). 

samplePercent: 10

textLabel

For training either with language model objectives or text/token classification, your data needs to point to the text. The column with this information is specified with textLabel. The default is "text" which means your text data for training should be labeled with a column with that name. 

textLabel: text

pairLabel

For text classification, you can specify two sentences for tasks like natural language inference and paraphrase detection. You can specify the column with the second sentence using pairLabel. The default is "pair". The example below shows how to specify the correct columns for glue's mnli task. 

textLabel: premise
pairLabel: hypothesis

tokensLabel

For token classification, the dataset must also provide the tokens (e.g., the words). You can specify the column with this data with tokensLabel. The default is "tokens". 

tokensLabel: tokens

tagsLabel

For token classification, the dataset must also provide the per-token tags (e.g., the named-entity tags). You can specify the column with this data with tagsLabel. The default is "tags". 

tagsLabel: tags

Extra Training Settings

You can specify the following additional training settings using the config file (lightly adapted from HuggingFace's Trainer arguments and Trainer class):

modelfpath: The output directory where the model is saved
epochs: Number of training epochs. The default is 2. 
eval_strategy: The evaluation strategy to use ('no' | 'steps'
                | 'epoch'). The default is 'epoch'.
eval_steps: Number of update steps between two evaluations.
                    The default is 500. 
batchSize: The per-device batch size for train/eval. The
                    default is 8. 
learning_rate: The initial learning rate for AdamW. Default
                        is 5e-5. 
weight_decay: Weight decay. The default is 0.01.
save_strategy: The checkpoint save strategy to use ('no'
                | 'steps' | 'epoch'). The default is 'epoch'.
save_steps: Number of update steps between saves. Default is
                    500.
load_best_model_at_end: Whether or not to load the best model
                at the end of training. If True, the best model will
                be saved as the model at the end. Default is False.
maskProbability: The rate of dynamic masking for masked
                language modeling. The default is 0.15.

Additional Config Settings

There are further parameters that can be specified, detailed below.

device

You can specify the device you want to run the model on with device: 

device: mps

The options are best, mps, cuda, or cpu (specific devices can also be specified). The default is best, which prioritizes mps and then cpu. (You must specify cuda or similar to use GPUs).

precision

You can control the memory requirements for your experiments with precision, which controls the precision of the loaded model (if applicable): 

precision: 16bit

The options are full, 16bit, 8bit, and 4bit. The default behavior is full, which loads the model without changing its precision. Selecting 16bit with train will train a lower precision model (note: you need to use a GPU for this).

PLL_type

PLL_type: original

For masked language models, the predictability measures are gathered by iteratively masking each token in the input. Following, Kauf and Ivanova (2023), this 'pseudo-likelihood' can be determined via different masking schemes. Presently we have two options, original which simply masks each token in the input one by one, following Salazar et al. (2020), and within_word_l2r which handles words that are subworded by masking each subword token to the right of the word being predicted. The default is within_word_l2r as Kauf and Ivanova find it performs better.

id2label

For classification models, it can be helpful to specify a mapping from model labels to identifiable. You can do this by specifying the mappings, as below: 

id2label:
    1: Positive
    0: Negative

This maps the output index 1 to Positive and 0 to Negative. The default behavior is to use the models id2label attribute.

tokenizers

You can specify the specific tokenizers you want to use with tokenizers:

models: 
        hf_causal_model:
            - gpt2-medium
tokenizers:
        hf_tokenizer: 
            - gpt2

At the moment, the only supported tokenizers are those using HuggingFace's AutoTokenizer class. Note that the tokenizers and models are aligned lazily. The assumption is that for each model listed in models, the tokenizer is in the same position in the tokenizer list. For example, 

models: 
        hf_causal_model:
            - gpt2
        hf_masked_model:
            - bert-base-uncased
tokenizers:
        hf_tokenizer: 
            - bert-base-uncased
            - gpt2

Loads GPT2 with BERT's tokenizer and BERT with GPT2's tokenizer. The default behavior is to load the version of the tokenizer associated with the model. The default behavior is strongly encouraged.

doLower

The input text can be lowercased with doLower, which takes a boolean. If set to True the text is lowercased (with special tokens preserved). The default behavior is False. Note this does not override the tokenizer's default behavior. For example, if you do not set doLower to True but use bert-base-uncased, the text will still be lowercased. 

doLower: True

addPrefixSpace

Some tokenizers include white space in their tokens (e.g., Byte-based tokenizers as in GPT2). For these tokenizers, the presence of a initial space affects tokenization. You can explicitly set this behavior with addPrefixSpace, which will add a prefix space if set to True. The default behavior is to not add a space (i.e., False), following other libraries like minicons. 

addPrefixSpace: True

addPadToken

For batched input, it is important to have a pad token set. This can be done with addPadToken. You can either put True or eos_token if you want to use the model's eos token or give a specific token (e.g., PAD). Note, only a word with a single token id (i.e., a word not sub-worded) can be used as a pad token. An error will be thrown if this is not the case. Note, this will not override the model's pad token if one already exists. The default behavior is to use eos, if no pad token exists. If none is provided and their is no eos token in the model, an error may be thrown during tokenization. 

addPadToken: True

batchSize

In experiments, you can control the batch size of the model with batchSize. The default size is 1 in evaluate mode and 16 in train mode.

batchSize: 1

verbose

In running experiments, you can control verbosity with verbose. When set to True, it prints some more information to the screen. The default behavior is True. 

verbose: True

Citation

Please cite our NLP Scholar paper: 

@inproceedings{prasad-davis-2024-training-nlp,
  title = {Training an {NLP} Scholar at a Small Liberal Arts College: A Backwards Designed Course Proposal},
  author = {Prasad, Grusha and Davis, Forrest},
  editor = {Al-azzawi, Sana and Biester, Laura and Kov{\'a}cs, Gy{\"o}rgy and Marasovi{\'c}, Ana and Mathur, Leena and Mieskes, Margot and Weissweiler, Leonie},
  booktitle = {Proceedings of the Sixth Workshop on Teaching NLP},
  month = aug,
  year = {2024},
  address = {Bangkok, Thailand},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2024.teachingnlp-1.16},
  pages = {105--118},
}

This wouldn't work without HuggingFace! Please cite them too: 

@inproceedings{wolf-etal-2020-transformers,
    title = "Transformers: State-of-the-Art Natural Language Processing",
    author = "Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
    month = oct,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-demos.6",
    pages = "38--45"
}

@inproceedings{lhoest-etal-2021-datasets,
    title = "Datasets: A Community Library for Natural Language Processing",
    author = "Lhoest, Quentin  and
      Villanova del Moral, Albert  and
      Jernite, Yacine  and
      Thakur, Abhishek  and
      von Platen, Patrick  and
      Patil, Suraj  and
      Chaumond, Julien  and
      Drame, Mariama  and
      Plu, Julien  and
      Tunstall, Lewis  and
      Davison, Joe  and
      {\v{S}}a{\v{s}}ko, Mario  and
      Chhablani, Gunjan  and
      Malik, Bhavitvya  and
      Brandeis, Simon  and
      Le Scao, Teven  and
      Sanh, Victor  and
      Xu, Canwen  and
      Patry, Nicolas  and
      McMillan-Major, Angelina  and
      Schmid, Philipp  and
      Gugger, Sylvain  and
      Delangue, Cl{\'e}ment  and
      Matussi{\`e}re, Th{\'e}o  and
      Debut, Lysandre  and
      Bekman, Stas  and
      Cistac, Pierric  and
      Goehringer, Thibault  and
      Mustar, Victor  and
      Lagunas, Fran{\c{c}}ois  and
      Rush, Alexander  and
      Wolf, Thomas",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-demo.21",
    pages = "175--184",
    abstract = "The scale, variety, and quantity of publicly-available NLP datasets has grown rapidly as researchers propose new tasks, larger models, and novel benchmarks. Datasets is a community library for contemporary NLP designed to support this ecosystem. Datasets aims to standardize end-user interfaces, versioning, and documentation, while providing a lightweight front-end that behaves similarly for small datasets as for internet-scale corpora. The design of the library incorporates a distributed, community-driven approach to adding datasets and documenting usage. After a year of development, the library now includes more than 650 unique datasets, has more than 250 contributors, and has helped support a variety of novel cross-dataset research projects and shared tasks. The library is available at https://github.com/huggingface/datasets.",
    eprint={2109.02846},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
}

If you use Kauf and Ivanova's PPL scoring technique, please cite them: 

@inproceedings{kauf-ivanova-2023-better,
    title = "A Better Way to Do Masked Language Model Scoring",
    author = "Kauf, Carina  and
      Ivanova, Anna",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-short.80",
    doi = "10.18653/v1/2023.acl-short.80",
    pages = "925--935",
    abstract = "Estimating the log-likelihood of a given sentence under an autoregressive language model is straightforward: one can simply apply the chain rule and sum the log-likelihood values for each successive token. However, for masked language models (MLMs), there is no direct way to estimate the log-likelihood of a sentence. To address this issue, Salazar et al. (2020) propose to estimate sentence pseudo-log-likelihood (PLL) scores, computed by successively masking each sentence token, retrieving its score using the rest of the sentence as context, and summing the resulting values. Here, we demonstrate that the original PLL method yields inflated scores for out-of-vocabulary words and propose an adapted metric, in which we mask not only the target token, but also all within-word tokens to the right of the target. We show that our adapted metric (PLL-word-l2r) outperforms both the original PLL metric and a PLL metric in which all within-word tokens are masked. In particular, it better satisfies theoretical desiderata and better correlates with scores from autoregressive models. Finally, we show that the choice of metric affects even tightly controlled, minimal pair evaluation benchmarks (such as BLiMP), underscoring the importance of selecting an appropriate scoring metric for evaluating MLM properties.",
}

Other Resources

If you like this toolkit and want more tools or different tools, check out the following repositories: