license: apache-2.0
language: en
tags:
- deberta-v3-base
- deberta-v3
- deberta
- text-classification
- nli
- natural-language-inference
- multitask
- multi-task
- pipeline
- extreme-multi-task
- extreme-mtl
- tasksource
- zero-shot
- rlhf
model-index:
- name: deberta-v3-base-tasksource-nli
results:
- task:
type: text-classification
name: Text Classification
dataset:
name: glue
type: glue
config: rte
split: validation
metrics:
- type: accuracy
value: 0.89
datasets:
- glue
- super_glue
- anli
- tasksource/babi_nli
- sick
- snli
- scitail
- OpenAssistant/oasst1
- universal_dependencies
- hans
- qbao775/PARARULE-Plus
- alisawuffles/WANLI
- metaeval/recast
- sileod/probability_words_nli
- joey234/nan-nli
- pietrolesci/nli_fever
- pietrolesci/breaking_nli
- pietrolesci/conj_nli
- pietrolesci/fracas
- pietrolesci/dialogue_nli
- pietrolesci/mpe
- pietrolesci/dnc
- pietrolesci/gpt3_nli
- pietrolesci/recast_white
- pietrolesci/joci
- martn-nguyen/contrast_nli
- pietrolesci/robust_nli
- pietrolesci/robust_nli_is_sd
- pietrolesci/robust_nli_li_ts
- pietrolesci/gen_debiased_nli
- pietrolesci/add_one_rte
- metaeval/imppres
- pietrolesci/glue_diagnostics
- hlgd
- PolyAI/banking77
- paws
- quora
- medical_questions_pairs
- conll2003
- Anthropic/hh-rlhf
- Anthropic/model-written-evals
- truthful_qa
- nightingal3/fig-qa
- tasksource/bigbench
- blimp
- cos_e
- cosmos_qa
- dream
- openbookqa
- qasc
- quartz
- quail
- head_qa
- sciq
- social_i_qa
- wiki_hop
- wiqa
- piqa
- hellaswag
- pkavumba/balanced-copa
- 12ml/e-CARE
- art
- tasksource/mmlu
- winogrande
- codah
- ai2_arc
- definite_pronoun_resolution
- swag
- math_qa
- metaeval/utilitarianism
- mteb/amazon_counterfactual
- SetFit/insincere-questions
- SetFit/toxic_conversations
- turingbench/TuringBench
- trec
- tals/vitaminc
- hope_edi
- strombergnlp/rumoureval_2019
- ethos
- tweet_eval
- discovery
- pragmeval
- silicone
- lex_glue
- papluca/language-identification
- imdb
- rotten_tomatoes
- ag_news
- yelp_review_full
- financial_phrasebank
- poem_sentiment
- dbpedia_14
- amazon_polarity
- app_reviews
- hate_speech18
- sms_spam
- humicroedit
- snips_built_in_intents
- banking77
- hate_speech_offensive
- yahoo_answers_topics
- pacovaldez/stackoverflow-questions
- zapsdcn/hyperpartisan_news
- zapsdcn/sciie
- zapsdcn/citation_intent
- go_emotions
- scicite
- liar
- relbert/lexical_relation_classification
- metaeval/linguisticprobing
- tasksource/crowdflower
- metaeval/ethics
- emo
- google_wellformed_query
- tweets_hate_speech_detection
- has_part
- wnut_17
- ncbi_disease
- acronym_identification
- jnlpba
- species_800
- SpeedOfMagic/ontonotes_english
- blog_authorship_corpus
- launch/open_question_type
- health_fact
- commonsense_qa
- mc_taco
- ade_corpus_v2
- prajjwal1/discosense
- circa
- YaHi/EffectiveFeedbackStudentWriting
- Ericwang/promptSentiment
- Ericwang/promptNLI
- Ericwang/promptSpoke
- Ericwang/promptProficiency
- Ericwang/promptGrammar
- Ericwang/promptCoherence
- PiC/phrase_similarity
- copenlu/scientific-exaggeration-detection
- quarel
- mwong/fever-evidence-related
- numer_sense
- dynabench/dynasent
- raquiba/Sarcasm_News_Headline
- sem_eval_2010_task_8
- demo-org/auditor_review
- medmcqa
- aqua_rat
- RuyuanWan/Dynasent_Disagreement
- RuyuanWan/Politeness_Disagreement
- RuyuanWan/SBIC_Disagreement
- RuyuanWan/SChem_Disagreement
- RuyuanWan/Dilemmas_Disagreement
- lucasmccabe/logiqa
- wiki_qa
- metaeval/cycic_classification
- metaeval/cycic_multiplechoice
- metaeval/sts-companion
- metaeval/commonsense_qa_2.0
- metaeval/lingnli
- metaeval/monotonicity-entailment
- metaeval/arct
- metaeval/scinli
- metaeval/naturallogic
- onestop_qa
- demelin/moral_stories
- corypaik/prost
- aps/dynahate
- metaeval/syntactic-augmentation-nli
- metaeval/autotnli
- lasha-nlp/CONDAQA
- openai/webgpt_comparisons
- Dahoas/synthetic-instruct-gptj-pairwise
- metaeval/scruples
- metaeval/wouldyourather
- sileod/attempto-nli
- metaeval/defeasible-nli
- metaeval/help-nli
- metaeval/nli-veridicality-transitivity
- metaeval/natural-language-satisfiability
- metaeval/lonli
- metaeval/dadc-limit-nli
- ColumbiaNLP/FLUTE
- metaeval/strategy-qa
- openai/summarize_from_feedback
- metaeval/folio
- metaeval/tomi-nli
- metaeval/avicenna
- stanfordnlp/SHP
- GBaker/MedQA-USMLE-4-options-hf
- sileod/wikimedqa
- declare-lab/cicero
- amydeng2000/CREAK
- metaeval/mutual
- inverse-scaling/NeQA
- inverse-scaling/quote-repetition
- inverse-scaling/redefine-math
- metaeval/puzzte
- metaeval/implicatures
- race
- metaeval/spartqa-yn
- metaeval/spartqa-mchoice
- metaeval/temporal-nli
- metaeval/ScienceQA_text_only
- AndyChiang/cloth
- metaeval/logiqa-2.0-nli
- tasksource/oasst1_dense_flat
- metaeval/boolq-natural-perturbations
- metaeval/path-naturalness-prediction
- riddle_sense
- Jiangjie/ekar_english
- metaeval/implicit-hate-stg1
- metaeval/chaos-mnli-ambiguity
- IlyaGusev/headline_cause
- metaeval/race-c
- metaeval/equate
- metaeval/ambient
- AndyChiang/dgen
- metaeval/clcd-english
- civil_comments
- metaeval/acceptability-prediction
- maximedb/twentyquestions
- metaeval/counterfactually-augmented-snli
- tasksource/I2D2
- sileod/mindgames
- metaeval/counterfactually-augmented-imdb
- metaeval/cnli
- metaeval/reclor
- tasksource/oasst1_pairwise_rlhf_reward
metrics:
- accuracy
library_name: transformers
pipeline_tag: zero-shot-classification
Model Card for DeBERTa-v3-base-tasksource-nli
This is DeBERTa-v3-base fine-tuned with multi-task learning on 560 tasks of the tasksource collection This checkpoint has strong zero-shot validation performance on many tasks (e.g. 70% on WNLI), and can be used for zero-shot NLI pipeline (similar to bart-mnli but better). You can also load other tasks (see next paragraph) or further fine-tune the encoder for new task (classification, token classification or multiple-choice).
Tasksource-adapters: 1 line access to more than 500 tasks
!pip install tasknet tasksource
import tasknet as tn
pipe = tn.load_pipeline('sileod/deberta-v3-base-tasksource-nli','glue/sst2') # works for 500+ tasksource tasks
pipe(['That movie was great !', 'Awful movie.'])
# [{'label': 'positive', 'score': 0.9956}, {'label': 'negative', 'score': 0.9967}]
The list of tasks is available in model config.json.
Evaluation
This model ranked 1st among all models with the microsoft/deberta-v3-base architecture according to the IBM model recycling evaluation. Results: 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.
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
Software and training details
https://github.com/sileod/tasksource/
https://github.com/sileod/tasknet/
Training code: https://colab.research.google.com/drive/1iB4Oxl9_B5W3ZDzXoWJN-olUbqLBxgQS?usp=sharing
Training took 7 days on RTX6000 24GB gpu.
This is the shared model with the MNLI classifier on top. Datasets including bigbench, Anthropic rlhf, anli... alongside many NLI and classification. 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 64k. The model was trained for 45k steps with a batch size of 384, and a peak learning rate of 2e-5.
Citation
More details on this article:
@article{sileo2023tasksource,
title={tasksource: Structured Dataset Preprocessing Annotations for Frictionless Extreme Multi-Task Learning and Evaluation},
author={Sileo, Damien},
url= {https://arxiv.org/abs/2301.05948},
journal={arXiv preprint arXiv:2301.05948},
year={2023}
}