9 ноября 2021 г. 10:19
Команда Yandex
Ссылка на модель https://habr.com/ru/company/yandex/blog/588214/
| Датасет | Результат | Метрика |
|---|---|---|
| LiDiRus | 0,364 | Кор, коэффициент Мэтью |
| RCB | 0,357 / 0,479 | F1/Точность |
| PARus | 0,834 | Точность |
| MuSeRC | 0,892 / 0,707 | F1a/Em |
| TERRa | 0,841 | Точность |
| RUSSE | 0,71 | Точность |
| RWSD | 0,669 | Точность |
| DaNetQA | 0,85 | Точность |
| RuCoS | 0,92 / 0,916 | F1/EM |
Another representative of YaLM — family of autoregressive transformers pretrained on large corpus of filtered runet, Russian Wikipedia, RDT and Taiga datasets, social media dialogs and articles from Yandex.News. The model has 28 layers (3.3 billion parameters). All tasks were solved with P-tuning: instead of using hand-crafted promts or complete fine-tuning, we optimize only 40k parameters on RSG training sets.
| Категория | Результат |
|---|---|
| LOGIC | 0,2220002391136312 |
| KNOWLEDGE | 0,3731079163231701 |
| PREDICATE-ARGUMENT STRUCTURE | 0,3379586695984027 |
| LEXICAL SEMANTICS | 0,4914961200052112 |
| Lexical Semantics - Lexical Entailment | 0,44614865235538076 |
|---|---|
| Lexical Semantics - Morphological Negation | 0,5619514869490163 |
| Lexical Semantics - Factivity | 0,45098039215686275 |
| Lexical Semantics - Symmetry/Collectivity | 0,570544330734548 |
| Lexical Semantics - Redundancy | 0,27348301713730944 |
| Lexical Semantics - Named Entities | 0,35355339059327373 |
| Lexical Semantics - Quantifiers | 0,32256184896566575 |
| Predicate-Argument Structure Core Args | 0,46312262138612026 |
| Predicate-Argument Structure Prepositional Phrases | 0,6257214261204825 |
| Predicate-Argument Structure Ellipsis/Implicits | 0,3490449235764129 |
| Predicate-Argument Structure Anaphora/Coreference | -0,004901960784313725 |
| Predicate-Argument Structure Active/Passive | 0,12565617248750865 |
| Predicate-Argument Structure Nominalization | 0,5017348819226064 |
| Predicate-Argument Structure Genitives/Partitives | 0,7637626158259734 |
| Predicate-Argument Structure Datives | 0,2058790548922549 |
| Predicate-Argument Structure Relative Clauses | 0,24370871833797697 |
| Predicate-Argument Structure Coordination Scopes | 0,366007208697342 |
| Predicate-Argument Structure Intersectivity | 0,2370100583365928 |
| Predicate-Argument Structure Restrictivity | 0,3619613829965134 |
| Logic Negation | 0,27900096804546626 |
| Logic Double Negation | 0,30151134457776363 |
| Logic Interval/Numbers | -0,27124713338229184 |
| Logic Conjuction | 0,6 |
| Logic Disjunction | 0,27608872259967787 |
| Logic Conditionals | 0,07100716024967263 |
| Logic Universal | 0,6446583712203042 |
| Logic Existential | 0,17910620335162064 |
| Logic Temporal | -0,15749883157896472 |
| Logic Upward Monotone | 0,7190325425674107 |
| Logic Downward Monotone | -0,22171945701357465 |
| Logic Non-Monotonic | 0,05572782125753529 |
| Knowledge Common Sense | 0,36452669663330906 |
| Knowledge World Knowledge | 0,36188997993038646 |
| Датасет | Speed | RAM |
|---|---|---|
| LiDiRus | - | - |
| RCB | - | - |
| PARus | - | - |
| MuSeRC | - | - |
| TERRa | - | - |
| RUSSE | - | - |
| RWSD | - | - |
| DaNetQA | - | - |
| RuCoS | - | - |