godofnothing/lm-kernel-ft

Based on https://github.com/princeton-nlp/LM-Kernel-FT

A Kernel-Based View of Language Model Fine-Tuning

This is the implementation for the paper A Kernel-Based View of Language Model Fine-tuning and can be used to compute kernel approximations for the fine-tuning of pre-trained language models.

We extend the LM-BFF repository and add a new "kernel trainer" powered by functorch to compute empirical-NTK kernel matrices using the SGD, SignGD or Asymmetric-SignGD kernel formulas. We also provide our pre-computed kernels for download to facilitate further analysis.

Installation

Please install all the dependency packages by using the following command:

pip install -r requirements.txt

We updated the LM-BFF code to work with a newer version of HuggingFace transformers and additionally require functorch. If you would like to run LoRA fine-tuning, install the LoRA version of the transformers library (see here) and add the flags --apply_lora --lora_alpha .... --lora_r ... .

NOTE: Different versions of some packages (pytorch, numpy, transformers) may cause minor variations in kernels and results.

Prepare the data

Please run the following commands to download and prepare the data:

( cd data; bash download_dataset.sh )

for K in 16 64 512; do
    # Generate k-shot splits for seeds 13,21,42,87,100 with a maximum of 1k test examples in data/k-shot-1k-test,
    # where k is the number of training/validation examples per label
    python tools/generate_k_shot_data.py --mode k-shot-1k-test --k $K
done

This follows LM-BFF, but download_dataset.sh additionally rebalances the cr dataset and uses the GLUE version of the SST-2 dataset. Additionally k-shot-1k-test limits test datasets to 1k examples for faster evaluation.

NOTE: During training, the model will generate/load cache files in the data folder. If your data have changed, make sure to clean all the cache files (starting with "cache").

Run the code

To easily run our experiments, you can use run_fewshot.sh:

TAG=kernel-prompting TRAINER=kernel TASK=SST-2 SEED=42 MODEL=roberta-base bash run_fewshot.sh

The templates and label word mappings are already defined, so you only need to set hyper-parameters and TAG (you can use whatever tag you want and it just makes finding results easier). See run_fewshot.sh for more options. Besides, you can easily add extra arguments:

NUM_GPU=4 TAG=kernel-prompting TRAINER=kernel TASK=SST-2 SEED=42 MODEL=roberta-base bash run_fewshot.sh \
    --kernel_formula signgd --kernel_solver logistic  --per_device_train_batch_size 2 --per_device_eval_batch_size 4

This splits the kernel computation across 4 GPUs and uses the SignGD kernel formula and a logistic kernel solver (the default is least-squares regression) and uses batch sizes 2 and 4 along the two axes of the kernel matrices respectively.

For more advanced use cases, such as how to aggregate results over multiple runs, zero-shot experiments or writing your own prompt formats, we refer to the README in the LM-BFF repo. Note that we deleted some tools to do automatic prompt and label search that are unrelated to our paper.

Download our pre-computed kernels

Here are the links for downloading our pre-computed kernels:

• SGD kernels
• SignGD kernels
• Asymmetric-SignGD kernels

The provided kernels were computed with RoBERTa-base for 12 datasets (SST-2, MR, CR, MPQA, Subj, TREC, MNLI, SNLI, QNLI, RTE, MRPC, QQP) over 5 seeds on both 16-shot and 64-shot datasets, where k-shot is the number of training/validation examples per label. The SGD kernels also include 6 datasets (SST-2, MR, CR, QNLI, RTE, QQP) for 512-shot datasets.

For each task and data split, we include separate files for training, development, test kernel matrices. Each file can be read using torch.load and contains a tuple of (kernel matrix, labels), and the kernel matrix has the shape of [training examples, training logits, X examples, X logits], where X dataset is given by the file name (train, dev or test).

Bugs and questions?

If you have any questions related to the code or the paper, feel free to email Alexander and Sadhika ({awettig,smalladi}@cs.princeton.edu). If you encounter a problem or bug when using the code, you can also open an issue.

Citation

Please cite our work if you make use of our code or our pre-computed kernels in your work:

@article{malladi2022kernel,
      title={A Kernel-Based View of Language Model Fine-Tuning},
      author={Malladi, Sadhika and Wettig, Alexander and Yu, Dingli and Chen, Danqi and Arora, Sanjeev},
      journal={arXiv preprint arXiv:2210.05643},
      year={2022}
}