This repo provides the source code for our method

• CET : a fine-tuning method for preserving commonsense knowledge from Pre-trained LMs

Besides, we also provide re-implementation of the following methods for a fair comparison:

• Vanilla fine-tuning: fine-tuning has been proven to be a simple and effective method of adapting large PLMs to downstream tasks.
• BSS: BSS focuses on mitigating negative transfer by penalizing the small singular values of the feature matrix.
• ChildTune-F&ChildTune-D: ChildTune-F&ChildTune-D update a subset of parameters (called child network) of large PLMs in the backward process. ChildTune-D utilizes the Fisher Information Matrix estimated by the pre-trained model to determine the child network. ChildTune-F uses Bernoulli distribution to determine the child network.
• Mixout: Mixout randomly mixes the parameters of the pre-trained and the fine-tuned model to regularize the fine-tuning process.
• NoisyTune: NoisyTune adds uniform noises to the parameter of the pre-trained model based on their standard deviations.
• R3F: R3F alleviates representational collapse by introducing parametric noise. R3F generates noise from either a normal or uniform distribution.
• RecAdam: RecAdam optimizes a multi-task objective and utilize an annealing coefficient to gradually shift the objective from pre-training to downstream tasks.
• ReInit: Zhang et. al. verified that transferring the top pre-trained layers slows down learning and hurts performance. ReInit re-initializes the top layers of PLMs when adapting to new tasks.

• Overview of the directory

• config/ : the directory of configurations for each models
• modeling/ : modeling the pre-trained models
• utils/ : implementing different fine-tuning methods
• main.py : the python file to be executed

./
├── config
│   ├── arc_challenge
│   ├── arc_easy
│   ├── csqa
│   ├── siqa
│   ├── obqa
│   ├── piqa
│   ├── qasc
│   └── default.yaml
├── main.py
├── modeling
│   └── bert_models.py
└── utils
    ├── data_utils.py
    ├── layers.py
    ├── logger_utils.py
    ├── mixout_utils.py
    ├── optimization_utils.py
    ├── other_utils.py
    ├── parser_utils.py
    ├── split_data_uils.py
    └── tokenization_utils.py

Reference environment settings:

python             3.9.13
torch              1.13.1+cu117
transformers       4.30.2

You can specify the path for storing datasets and saving logs in ./utils/parser_utils.py.

For example, the default paths for loading the OBQA dataset are ./data/obqa/official/train.jsonl, ./data/obqa/official/dev.jsonl, and ./data/obqa/official/test.jsonl.

The .jsonl files have the format as follows:

{"id": "7-980", "question": {"stem": "The sun is responsible for", "choices": [{"text": "puppies learning new tricks", "label": "A"}, {"text": "children growing up and getting old", "label": "B"}, {"text": "flowers wilting in a vase", "label": "C"}, {"text": "plants sprouting, blooming and wilting", "label": "D"}]}, "answerKey": "D"}
{"id": "7-584", "question": {"stem": "When standing miles away from Mount Rushmore", "choices": [{"text": "the mountains seem very close", "label": "A"}, {"text": "the mountains are boring", "label": "B"}, {"text": "the mountains look the same as from up close", "label": "C"}, {"text": "the mountains seem smaller than in photographs", "label": "D"}]}, "answerKey": "D"}
{"id": "7-870", "question": {"stem": "When food is reduced in the stomach", "choices": [{"text": "the mind needs time to digest", "label": "A"}, {"text": "take a second to digest what I said", "label": "B"}, {"text": "nutrients are being deconstructed", "label": "C"}, {"text": "reader's digest is a body of works", "label": "D"}]}, "answerKey": "C"}
{"id": "7-321", "question": {"stem": "Stars are", "choices": [{"text": "warm lights that float", "label": "A"}, {"text": "made out of nitrate", "label": "B"}, {"text": "great balls of gas burning billions of miles away", "label": "C"}, {"text": "lights in the sky", "label": "D"}]}, "answerKey": "C"}
...

You can download the data we used in this link, or you can use your own datasets with the same format as above.

Specify your configurations (e.g., ./config/obqa/roberta_large_baseline.yaml) and run the following command

python3 main.py --config {selected_yaml_configuration_path} --run_id {your_experiment_id}

Then, the results as well as the model checkpoint will be saved automatically in the directory ./save_models/ (in default).

If you find the code useful, please consider citing this work

Junhao Zheng, Qianli Ma, Shengjie Qiu, Yue Wu, Peitian Ma, Junlong Liu, Huawen Feng, Xichen Shang, and Haibin Chen. 2023. Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9155–9173, Toronto, Canada. Association for Computational Linguistics.