license: apache-2.0
language: en
tags:
- deberta-v3-base
- text-classification
- nli
- natural-language-inference
- multitask
- extreme-mtl
- deberta-v3-base
pipeline_tag: zero-shot-classification
datasets:
- hellaswag
- ag_news
- pietrolesci/nli_fever
- numer_sense
- go_emotions
- Ericwang/promptProficiency
- poem_sentiment
- pietrolesci/robust_nli_is_sd
- sileod/probability_words_nli
- social_i_qa
- trec
- pietrolesci/gen_debiased_nli
- snips_built_in_intents
- metaeval/imppres
- metaeval/crowdflower
- tals/vitaminc
- dream
- metaeval/babi_nli
- Ericwang/promptSpoke
- metaeval/ethics
- art
- ai2_arc
- discovery
- Ericwang/promptGrammar
- code_x_glue_cc_clone_detection_big_clone_bench
- prajjwal1/discosense
- pietrolesci/joci
- Anthropic/model-written-evals
- utilitarianism
- emo
- tweets_hate_speech_detection
- piqa
- blog_authorship_corpus
- SpeedOfMagic/ontonotes_english
- circa
- app_reviews
- anli
- Ericwang/promptSentiment
- codah
- definite_pronoun_resolution
- health_fact
- tweet_eval
- hate_speech18
- glue
- hendrycks_test
- paws
- bigbench
- hate_speech_offensive
- blimp
- sick
- turingbench/TuringBench
- martn-nguyen/contrast_nli
- Anthropic/hh-rlhf
- openbookqa
- species_800
- alisawuffles/WANLI
- ethos
- pietrolesci/mpe
- wiki_hop
- pietrolesci/glue_diagnostics
- mc_taco
- quarel
- PiC/phrase_similarity
- strombergnlp/rumoureval_2019
- quail
- acronym_identification
- pietrolesci/robust_nli
- quora
- wnut_17
- dynabench/dynasent
- pietrolesci/gpt3_nli
- truthful_qa
- pietrolesci/add_one_rte
- pietrolesci/breaking_nli
- copenlu/scientific-exaggeration-detection
- medical_questions_pairs
- rotten_tomatoes
- scicite
- scitail
- pietrolesci/dialogue_nli
- code_x_glue_cc_defect_detection
- nightingal3/fig-qa
- pietrolesci/conj_nli
- liar
- sciq
- head_qa
- pietrolesci/dnc
- quartz
- wiqa
- code_x_glue_cc_code_refinement
- Ericwang/promptCoherence
- joey234/nan-nli
- hope_edi
- jnlpba
- yelp_review_full
- pietrolesci/recast_white
- swag
- banking77
- cosmos_qa
- financial_phrasebank
- hans
- pietrolesci/fracas
- math_qa
- conll2003
- qasc
- ncbi_disease
- mwong/fever-evidence-related
- YaHi/EffectiveFeedbackStudentWriting
- ade_corpus_v2
- amazon_polarity
- pietrolesci/robust_nli_li_ts
- super_glue
- adv_glue
- Ericwang/promptNLI
- cos_e
- launch/open_question_type
- lex_glue
- has_part
- pragmeval
- sem_eval_2010_task_8
- imdb
- humicroedit
- sms_spam
- dbpedia_14
- commonsense_qa
- hlgd
- snli
- hyperpartisan_news_detection
- google_wellformed_query
- raquiba/Sarcasm_News_Headline
- metaeval/recast
- winogrande
- relbert/lexical_relation_classification
- metaeval/linguisticprobing
metrics:
- accuracy
library_name: transformers
Model Card for DeBERTa-v3-base-tasksource-nli
DeBERTa model jointly fine-tuned on 444 tasks of the tasksource collection https://github.com/sileod/tasksource/ This is the model with the MNLI classifier on top. Its encoder was trained on many datasets including bigbench, Anthropic/hh-rlhf... alongside many NLI and classification tasks with a SequenceClassification heads while using only one shared encoder.
Each task had a specific CLS embedding, which is dropped 10% of the time to facilitate model use without it. All multiple-choice model used the same classification layers. For classification tasks, models shared weights if their labels matched. The number of examples per task was capped to 64. The model was trained for 20k steps with a batch size of 384, a peak learning rate of 2e-5.
You can fine-tune this model to use it for multiple-choice or any classification task (e.g. NLI) like any debertav2 model. This model has strong validation performance on many tasks (e.g. 70% on WNLI).
The list of tasks is available in tasks.md
code: https://colab.research.google.com/drive/1iB4Oxl9_B5W3ZDzXoWJN-olUbqLBxgQS?usp=sharing
Software
https://github.com/sileod/tasknet/ Training took 3 days on 24GB gpu.
Model Recycling
Evaluation on 36 datasets using sileod/deberta-v3-base_tasksource-420 as a base model yields average score of 80.45 in comparison to 79.04 by microsoft/deberta-v3-base.
An earlier (weaker) version model is ranked 1st among all tested models for the microsoft/deberta-v3-base architecture as of 10/01/2023 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 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
87.042 | 90.9 | 66.46 | 59.7188 | 85.5352 | 85.7143 | 87.0566 | 69 | 79.5333 | 91.6735 | 85.8 | 94.324 | 72.4902 | 90.2055 | 88.9706 | 63.9851 | 87.5 | 93.6299 | 91.7363 | 91.0882 | 84.4765 | 95.0688 | 56.9683 | 91.6654 | 98 | 91.2 | 46.814 | 84.3772 | 58.0471 | 81.25 | 85.2326 | 71.8821 | 69.4357 | 73.2394 | 74.0385 | 72.2 |
For more information, see: Model Recycling
Citation [optional]
BibTeX:
@misc{sileod23-tasksource,
author = {Sileo, Damien},
doi = {10.5281/zenodo.7473446},
month = {01},
title = {{tasksource: preprocessings for reproducibility and multitask-learning}},
url = {https://github.com/sileod/tasksource},
version = {1.5.0},
year = {2023}}