ColD Fusion model

Finetuned model that aims to be a great base model. It improves over RoBERTa base, trained on 35 datasets. Full details at this paper.

Paper Abstract:

Pretraining has been shown to scale well with compute, data size and data diversity. Multitask learning trains on a mixture of supervised datasets and produces improved performance compared to self-supervised pretraining. Until now, massively multitask learning required simultaneous access to all datasets in the mixture and heavy compute resources that are only available to well-resourced teams.

In this paper, we propose ColD Fusion, a method that provides the benefits of multitask learning but leverages distributed computation and requires limited communication and no sharing of data. Consequentially, ColD Fusion can create a synergistic loop, where finetuned models can be recycled to continually improve the pretrained model they are based on. We show that ColD Fusion yields comparable benefits to multitask pretraining by producing a model that (a) attains strong performance on all of the datasets it was multitask trained on and (b) is a better starting point for finetuning on unseen datasets. We find ColD Fusion outperforms RoBERTa and even previous multitask models. Specifically, when training and testing on 35 diverse datasets, ColD Fusion-based model outperforms RoBERTa by 2.45 points in average without any changes to the architecture.

How to use

Best way to use is to finetune on your own task, but you can also extract features directly. To get the features of a given text in PyTorch:

from transformers import RobertaTokenizer, RobertaModel
tokenizer = RobertaTokenizer.from_pretrained('ibm/ColD-Fusion')
model = RobertaModel.from_pretrained('ibm/ColD-Fusion')
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

and in TensorFlow:

from transformers import RobertaTokenizer, TFRobertaModel
tokenizer = RobertaTokenizer.from_pretrained('ibm/ColD-Fusion')
model = TFRobertaModel.from_pretrained('ibm/ColD-Fusion')
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)

Evaluation results

Model Recycling

Evaluation on 36 datasets using ibm/ColD-Fusion as a base model yields average score of 78.47 in comparison to 76.22 by roberta-base.

The model is ranked 1st among all tested models for the roberta-base architecture as of 21/12/2022 Results:

20_newsgroup	ag_news	amazon_reviews_multi	anli	boolq	cb	cola	copa	dbpedia	esnli	financial_phrasebank	imdb	isear	mnli	mrpc	multirc	poem_sentiment	qnli	qqp	rotten_tomatoes	rte	sst2	sst_5bins	stsb	trec_coarse	trec_fine	tweet_ev_emoji	tweet_ev_emotion	tweet_ev_hate	tweet_ev_irony	tweet_ev_offensive	tweet_ev_sentiment	wic	wnli	wsc	yahoo_answers
85.8205	89.8	66.26	51.9375	81.3761	87.5	83.3174	72	78.6333	91.1441	88.1	93.864	73.5332	87.2966	87.0098	63.717	85.5769	92.4034	91.113	91.8386	85.1986	95.4128	56.3801	91.2964	97	90.4	46.306	83.0401	54.4444	77.9337	85.9302	70.4331	68.652	47.8873	60.5769	71.8667

For more information, see: Model Recycling See full evaluation results of this model and many more here When fine-tuned on downstream tasks, this model achieves the following results:

BibTeX entry and citation info

@article{ColDFusion,
  author    = {Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, Leshem ChoshenYinhan Liu and},
  title     = {ColD Fusion: Collaborative Descent for Distributed Multitask Finetuning},
  journal   = {CoRR},
  volume    = {abs/2212.01378},
  year      = {2022},
  url       = {https://arxiv.org/abs/2212.01378},
  archivePrefix = {arXiv},
  eprint    = {2212.01378},
}