Этот репозиторий - развитие подхода к оценке моделей из поста Маленький и быстрый BERT для русского языка, эволюционировавшего в Рейтинг русскоязычных энкодеров предложений. Идея в том, чтобы понять, как хорошо разные модели превращают короткие тексты в осмысленные векторы.
Похожие проекты:
- RussianSuperGLUE: фокус на дообучаемых моделях
- MOROCCO: RussianSuperGLUE + оценка производительности, трудновоспроизводим
- RuSentEval: более академические/лингвистические задачи
- Статья от Вышки Popov et al, 2019: первая научная статья на эту тему, но маловато моделей и задач
- SentEvalRu и deepPavlovEval: два хороших, но давно не обновлявшихся бенчмарка.
Пример запуска метрик – в блокноте evaluation example.
Блокнот для воспроизведения лидерборда: v2021, v2023.
Лидерборд на HuggingFace Space.
Ранжирование моделей в по среднему качеству и производительности. Подсвечены Парето-оптимальные модели по каждому из критериев.
model | CPU | GPU | size | Mean S | Mean S+W | dim |
---|---|---|---|---|---|---|
sergeyzh/LaBSE-ru-turbo | 120.4 | 8.1 | 489.6 | 0.789 | 0.702 | 768 |
BAAI/bge-m3 | 523.4 | 22.5 | 2166.0 | 0.787 | 0.696 | 1024 |
intfloat/multilingual-e5-large-instruct | 501.5 | 25.71 | 2136.0 | 0.784 | 0.684 | 1024 |
intfloat/multilingual-e5-large | 506.8 | 30.8 | 2135.9389 | 0.78 | 0.686 | 1024 |
deepvk/USER-base | 33.1 | 12.2 | 473.2402 | 0.772 | 0.688 | 768 |
sentence-transformers/paraphrase-multilingual-mpnet-base-v2 | 20.5 | 19.9 | 1081.8485 | 0.762 | 768 | |
intfloat/multilingual-e5-base | 130.61 | 14.39 | 1061.0 | 0.761 | 0.669 | 768 |
sergeyzh/rubert-tiny-turbo | 5.5 | 3.3 | 111.4 | 0.749 | 0.667 | 312 |
intfloat/multilingual-e5-small | 40.86 | 12.09 | 449.0 | 0.742 | 0.645 | 384 |
symanto/sn-xlm-roberta-base-snli-mnli-anli-xnli | 20.2 | 16.5 | 1081.8474 | 0.739 | 768 | |
cointegrated/LaBSE-en-ru | 120.4 | 8.1 | 489.6621 | 0.739 | 0.668 | 768 |
sentence-transformers/LaBSE | 135.1 | 13.3 | 1796.5078 | 0.739 | 0.667 | 768 |
MUSE-3 | 200.1 | 30.7 | 303.0 | 0.736 | 512 | |
text-embedding-ada-002 | ? | 0.734 | 1536 | |||
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | 18.2 | 14.9 | 479.2547 | 0.734 | 384 | |
sentence-transformers/distiluse-base-multilingual-cased-v1 | 11.8 | 8.0 | 517.7452 | 0.722 | 512 | |
SONAR | ? | ? | 3060.0 | 0.721 | 1024 | |
facebook/nllb-200-distilled-600M | 252.3 | 15.9 | 1577.4828 | 0.709 | 0.64 | 1024 |
sentence-transformers/distiluse-base-multilingual-cased-v2 | 11.2 | 9.2 | 517.7453 | 0.708 | 512 | |
cointegrated/rubert-tiny2 | 5.5 | 3.3 | 111.3823 | 0.704 | 0.638 | 312 |
ai-forever/sbert_large_mt_nlu_ru | 504.5 | 29.7 | 1628.6539 | 0.703 | 0.626 | 1024 |
laser | 192.5 | 13.5 | 200.0 | 0.699 | 1024 | |
laser2 | 163.4 | 8.6 | 175.0 | 0.694 | 1024 | |
ai-forever/sbert_large_nlu_ru | 497.7 | 29.9 | 1628.6539 | 0.688 | 0.626 | 1024 |
clips/mfaq | 18.1 | 18.2 | 1081.8576 | 0.687 | 768 | |
cointegrated/rut5-base-paraphraser | 137.0 | 15.6 | 412.0015 | 0.685 | 0.634 | 768 |
DeepPavlov/rubert-base-cased-sentence | 128.4 | 13.2 | 678.5215 | 0.678 | 0.612 | 768 |
DeepPavlov/distilrubert-base-cased-conversational | 64.2 | 10.4 | 514.002 | 0.676 | 0.624 | 768 |
DeepPavlov/distilrubert-tiny-cased-conversational | 21.2 | 3.3 | 405.8292 | 0.67 | 0.616 | 768 |
cointegrated/rut5-base-multitask | 136.9 | 12.7 | 412.0015 | 0.668 | 0.623 | 768 |
ai-forever/ruRoberta-large | 512.3 | 25.5 | 1355.7162 | 0.666 | 0.609 | 1024 |
DeepPavlov/rubert-base-cased-conversational | 127.5 | 16.3 | 678.5215 | 0.653 | 0.606 | 768 |
deepvk/deberta-v1-base | 128.6 | 19.0 | 473.2402 | 0.653 | 0.591 | 768 |
cointegrated/rubert-tiny | 7.5 | 5.9 | 44.97 | 0.645 | 0.575 | 312 |
ai-forever/FRED-T5-large | 479.4 | 23.3 | 1372.9988 | 0.639 | 0.551 | 1024 |
inkoziev/sbert_synonymy | 6.9 | 4.2 | 111.3823 | 0.637 | 0.566 | 312 |
numind/NuNER-multilingual-v0.1 | 186.9 | 10 | 678.0 | 0.633 | 0.572 | 768 |
cointegrated/rubert-tiny-toxicity | 10 | 5.5 | 47.2 | 0.621 | 0.553 | 312 |
ft_geowac_full | 0.3 | 1910.0 | 0.617 | 0.55 | 300 | |
bert-base-multilingual-cased | 141.4 | 13.7 | 678.5215 | 0.614 | 0.565 | 768 |
ai-forever/ruT5-large | 489.6 | 20.2 | 1277.7571 | 0.61 | 0.578 | 1024 |
cointegrated/rut5-small | 37.6 | 8.6 | 111.3162 | 0.602 | 0.564 | 512 |
ft_geowac_21mb | 1.2 | 21.0 | 0.597 | 0.531 | 300 | |
inkoziev/sbert_pq | 7.4 | 4.2 | 111.3823 | 0.596 | 0.526 | 312 |
ai-forever/ruT5-base | 126.3 | 12.8 | 418.2325 | 0.571 | 0.544 | 768 |
hashing_1000_char | 0.5 | 1.0 | 0.557 | 0.464 | 1000 | |
cointegrated/rut5-base | 127.8 | 15.5 | 412.0014 | 0.554 | 0.53 | 768 |
hashing_300_char | 0.8 | 1.0 | 0.529 | 0.433 | 300 | |
hashing_1000 | 0.2 | 1.0 | 0.513 | 0.416 | 1000 | |
hashing_300 | 0.3 | 1.0 | 0.491 | 0.397 | 300 |
Ранжирование моделей по задачам. Подсвечены наилучшие модели по каждой из задач.
model | STS | PI | NLI | SA | TI | IA | IC | ICX | NE1 | NE2 |
---|---|---|---|---|---|---|---|---|---|---|
sergeyzh/LaBSE-ru-turbo | 0.86 | 0.75 | 0.49 | 0.81 | 0.97 | 0.81 | 0.82 | 0.80 | 0.30 | 0.40 |
BAAI/bge-m3 | 0.86 | 0.75 | 0.51 | 0.82 | 0.97 | 0.79 | 0.81 | 0.78 | 0.24 | 0.42 |
intfloat/multilingual-e5-large-instruct | 0.86 | 0.74 | 0.47 | 0.81 | 0.98 | 0.8 | 0.82 | 0.77 | 0.21 | 0.35 |
intfloat/multilingual-e5-large | 0.86 | 0.73 | 0.47 | 0.81 | 0.98 | 0.8 | 0.82 | 0.77 | 0.24 | 0.37 |
deepvk/USER-base | 0.85 | 0.74 | 0.48 | 0.81 | 0.99 | 0.81 | 0.8 | 0.7 | 0.29 | 0.41 |
sentence-transformers/paraphrase-multilingual-mpnet-base-v2 | 0.85 | 0.66 | 0.54 | 0.79 | 0.95 | 0.78 | 0.79 | 0.74 | ||
intfloat/multilingual-e5-base | 0.83 | 0.7 | 0.46 | 0.8 | 0.96 | 0.78 | 0.8 | 0.74 | 0.23 | 0.38 |
sergeyzh/rubert-tiny-turbo | 0.83 | 0.72 | 0.48 | 0.79 | 0.95 | 0.76 | 0.78 | 0.68 | 0.30 | 0.37 |
intfloat/multilingual-e5-small | 0.82 | 0.71 | 0.46 | 0.76 | 0.96 | 0.76 | 0.78 | 0.69 | 0.23 | 0.27 |
symanto/sn-xlm-roberta-base-snli-mnli-anli-xnli | 0.76 | 0.6 | 0.86 | 0.76 | 0.91 | 0.72 | 0.71 | 0.6 | ||
cointegrated/LaBSE-en-ru | 0.79 | 0.66 | 0.43 | 0.76 | 0.95 | 0.77 | 0.79 | 0.77 | 0.35 | 0.42 |
sentence-transformers/LaBSE | 0.79 | 0.66 | 0.43 | 0.76 | 0.95 | 0.77 | 0.79 | 0.76 | 0.35 | 0.41 |
MUSE-3 | 0.81 | 0.61 | 0.42 | 0.77 | 0.96 | 0.79 | 0.77 | 0.75 | ||
text-embedding-ada-002 | 0.78 | 0.66 | 0.44 | 0.77 | 0.96 | 0.77 | 0.75 | 0.73 | ||
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | 0.84 | 0.62 | 0.5 | 0.76 | 0.92 | 0.74 | 0.77 | 0.72 | ||
sentence-transformers/distiluse-base-multilingual-cased-v1 | 0.8 | 0.6 | 0.43 | 0.75 | 0.94 | 0.76 | 0.76 | 0.74 | ||
SONAR | 0.71 | 0.58 | 0.41 | 0.77 | 0.98 | 0.79 | 0.78 | 0.74 | ||
facebook/nllb-200-distilled-600M | 0.71 | 0.54 | 0.41 | 0.76 | 0.95 | 0.76 | 0.8 | 0.75 | 0.31 | 0.42 |
sentence-transformers/distiluse-base-multilingual-cased-v2 | 0.79 | 0.55 | 0.42 | 0.75 | 0.91 | 0.75 | 0.76 | 0.73 | ||
cointegrated/rubert-tiny2 | 0.75 | 0.65 | 0.42 | 0.74 | 0.94 | 0.75 | 0.76 | 0.64 | 0.36 | 0.39 |
ai-forever/sbert_large_mt_nlu_ru | 0.78 | 0.65 | 0.4 | 0.8 | 0.98 | 0.8 | 0.76 | 0.45 | 0.3 | 0.34 |
laser | 0.75 | 0.6 | 0.41 | 0.73 | 0.96 | 0.72 | 0.72 | 0.7 | ||
laser2 | 0.74 | 0.6 | 0.41 | 0.73 | 0.95 | 0.72 | 0.72 | 0.69 | ||
ai-forever/sbert_large_nlu_ru | 0.68 | 0.62 | 0.39 | 0.78 | 0.98 | 0.8 | 0.78 | 0.48 | 0.36 | 0.4 |
clips/mfaq | 0.63 | 0.59 | 0.35 | 0.79 | 0.95 | 0.74 | 0.76 | 0.69 | ||
cointegrated/rut5-base-paraphraser | 0.65 | 0.53 | 0.4 | 0.78 | 0.95 | 0.75 | 0.75 | 0.67 | 0.45 | 0.41 |
DeepPavlov/rubert-base-cased-sentence | 0.74 | 0.66 | 0.49 | 0.75 | 0.92 | 0.75 | 0.72 | 0.39 | 0.36 | 0.34 |
DeepPavlov/distilrubert-base-cased-conversational | 0.7 | 0.56 | 0.39 | 0.76 | 0.98 | 0.78 | 0.76 | 0.48 | 0.4 | 0.43 |
DeepPavlov/distilrubert-tiny-cased-conversational | 0.7 | 0.55 | 0.4 | 0.74 | 0.98 | 0.78 | 0.76 | 0.45 | 0.35 | 0.44 |
cointegrated/rut5-base-multitask | 0.65 | 0.54 | 0.38 | 0.76 | 0.95 | 0.75 | 0.72 | 0.59 | 0.47 | 0.41 |
ai-forever/ruRoberta-large | 0.7 | 0.6 | 0.35 | 0.78 | 0.98 | 0.8 | 0.78 | 0.32 | 0.3 | 0.46 |
DeepPavlov/rubert-base-cased-conversational | 0.68 | 0.52 | 0.38 | 0.73 | 0.98 | 0.78 | 0.75 | 0.42 | 0.41 | 0.43 |
deepvk/deberta-v1-base | 0.68 | 0.54 | 0.38 | 0.76 | 0.98 | 0.8 | 0.78 | 0.29 | 0.29 | 0.4 |
cointegrated/rubert-tiny | 0.66 | 0.53 | 0.4 | 0.71 | 0.89 | 0.68 | 0.7 | 0.58 | 0.24 | 0.34 |
ai-forever/FRED-T5-large | 0.62 | 0.44 | 0.37 | 0.78 | 0.98 | 0.81 | 0.67 | 0.45 | 0.25 | 0.15 |
inkoziev/sbert_synonymy | 0.69 | 0.49 | 0.41 | 0.71 | 0.91 | 0.72 | 0.69 | 0.47 | 0.32 | 0.24 |
numind/NuNER-multilingual-v0.1 | 0.67 | 0.53 | 0.4 | 0.71 | 0.89 | 0.72 | 0.7 | 0.46 | 0.32 | 0.34 |
cointegrated/rubert-tiny-toxicity | 0.57 | 0.44 | 0.37 | 0.68 | 1.0 | 0.78 | 0.7 | 0.43 | 0.24 | 0.32 |
ft_geowac_full | 0.69 | 0.53 | 0.37 | 0.72 | 0.97 | 0.76 | 0.66 | 0.26 | 0.22 | 0.34 |
bert-base-multilingual-cased | 0.66 | 0.53 | 0.37 | 0.7 | 0.89 | 0.7 | 0.69 | 0.38 | 0.36 | 0.38 |
ai-forever/ruT5-large | 0.51 | 0.39 | 0.35 | 0.77 | 0.97 | 0.79 | 0.72 | 0.38 | 0.46 | 0.44 |
cointegrated/rut5-small | 0.61 | 0.53 | 0.34 | 0.73 | 0.92 | 0.71 | 0.7 | 0.27 | 0.44 | 0.38 |
ft_geowac_21mb | 0.68 | 0.52 | 0.36 | 0.72 | 0.96 | 0.74 | 0.65 | 0.15 | 0.21 | 0.32 |
inkoziev/sbert_pq | 0.57 | 0.41 | 0.38 | 0.7 | 0.92 | 0.69 | 0.68 | 0.43 | 0.26 | 0.24 |
ai-forever/ruT5-base | 0.5 | 0.28 | 0.34 | 0.73 | 0.97 | 0.76 | 0.7 | 0.29 | 0.45 | 0.41 |
hashing_1000_char | 0.7 | 0.53 | 0.4 | 0.7 | 0.84 | 0.59 | 0.63 | 0.05 | 0.05 | 0.14 |
cointegrated/rut5-base | 0.44 | 0.28 | 0.33 | 0.74 | 0.92 | 0.75 | 0.58 | 0.39 | 0.48 | 0.39 |
hashing_300_char | 0.69 | 0.51 | 0.39 | 0.67 | 0.75 | 0.57 | 0.61 | 0.04 | 0.03 | 0.08 |
hashing_1000 | 0.63 | 0.49 | 0.39 | 0.66 | 0.77 | 0.55 | 0.57 | 0.05 | 0.02 | 0.04 |
hashing_300 | 0.61 | 0.48 | 0.4 | 0.64 | 0.71 | 0.54 | 0.5 | 0.05 | 0.02 | 0.02 |
- Semantic text similarity (STS) на основе переведённого датасета STS-B;
- Paraphrase identification (PI) на основе датасета paraphraser.ru;
- Natural language inference (NLI) на датасете XNLI;
- Sentiment analysis (SA) на данных SentiRuEval2016.
- Toxicity identification (TI) на датасете токсичных комментариев из OKMLCup;
- Inappropriateness identification (II) на датасете Сколтеха;
- Intent classification (IC) и её кросс-язычная версия ICX на датасете NLU-evaluation-data, который я автоматически перевёл на русский. В IC классификатор обучается на русских данных, а в ICX – на английских, а тестируется в обоих случаях на русских.
- Распознавание именованных сущностей на датасетах factRuEval-2016 (NE1) и RuDReC (NE2). Эти две задачи требуют получать эмбеддинги отдельных токенов, а не целых предложений; поэтому там участвуют не все модели.
- Август 2023 - обновил рейтинг:
- поправив ошибку в вычислении mean token embeddings
- добавил несколько моделей, включая нового лидера -
intfloat/multilingual-e5-large
- по просьбам трудящихся, добавил
text-embedding-ada-002
(размер и производительность указаны от балды)
- Лето 2022 - опубликовал первый рейтинг