Новый сабмит ruadapt LLaMA-2 7B LoRA

19 января 2024 г. 13:26

Команда RCC MSU

Ссылка на модель https://huggingface.co/rccmsu/ruadapt_llama2_7b_v0.1


Результат бейзлайна: 0,71

Датасет Результат Метрика
LiDiRus 0,417 Кор, коэффициент Мэтью
RCB 0,545 / 0,555 F1/Точность
PARus 0,756 Точность
MuSeRC 0,894 / 0,695 F1a/Em
TERRa 0,876 Точность
RUSSE 0,668 Точность
RWSD 0,708 Точность
DaNetQA 0,878 Точность
RuCoS 0,76 / 0,733 F1/EM
Описание модели:

Tikhomirov M., Chernyshev D. Impact of Tokenization on LLaMa Russian Adaptation //arXiv preprint arXiv:2312.02598. – 2023.


Описание параметров:

Tested using https://github.com/IlyaGusev/rulm repo { "trainer": { "evaluation_strategy": "steps", "per_device_train_batch_size": 4, "per_device_eval_batch_size": 4, "gradient_accumulation_steps": 32, "eval_steps": 50, "save_steps": 50, "logging_steps": 5, "learning_rate": 0.00025, "num_train_epochs": 3, "lr_scheduler_type": "cosine", "warmup_steps": 30, "fp16": true, "bf16": false, "torch_compile": false, "optim": "adamw_torch" }, "lora": { "r": 16, "lora_alpha": 16, "lora_dropout": 0.05, "bias": "none", "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"], "task_type": "CAUSAL_LM" }, "load_in_8bit": false, "only_target_loss": true, "mode": "chat", "templates_path": "internal_prompts/saiga_v2.json", "model_name": "llama2_7b_darulm_unigram_tie_2e_16_11_23", "model_type": "causal", "max_tokens_count": 1024 }

Диагностика: 0,417

Категория Результат
LOGIC 0,3472109173849461
KNOWLEDGE 0,3988089571509749
PREDICATE-ARGUMENT STRUCTURE 0,4107842841297232
LEXICAL SEMANTICS 0,4999636017616265
Lexical Semantics - Lexical Entailment 0,5648145000600431
Lexical Semantics - Morphological Negation 0,39477101697586137
Lexical Semantics - Factivity 0,4264014327112209
Lexical Semantics - Symmetry/Collectivity 0,31622776601683794
Lexical Semantics - Redundancy 0,17951965741648493
Lexical Semantics - Named Entities 0,612056372482123
Lexical Semantics - Quantifiers 0,3638714534805853
Predicate-Argument Structure Core Args 0,30962962962962964
Predicate-Argument Structure Prepositional Phrases 0,3943052207977212
Predicate-Argument Structure Ellipsis/Implicits 0,6205427141408237
Predicate-Argument Structure Anaphora/Coreference 0,4144368375833978
Predicate-Argument Structure Active/Passive 0,4083133966424866
Predicate-Argument Structure Nominalization 0,5444357229372963
Predicate-Argument Structure Genitives/Partitives 0,6813851438692469
Predicate-Argument Structure Datives 0,3563483225498992
Predicate-Argument Structure Relative Clauses 0,6407232755171874
Predicate-Argument Structure Coordination Scopes 0,14169568340005298
Predicate-Argument Structure Intersectivity 0,42208132696637884
Predicate-Argument Structure Restrictivity 0,36900620230837305
Logic Negation 0,473553991329486
Logic Double Negation 0,3892494720807615
Logic Interval/Numbers 0,08779776400125335
Logic Conjuction 0,33734954246999327
Logic Disjunction 0,4365575409204501
Logic Conditionals 0,3730235484764954
Logic Universal 0,3959441875175622
Logic Existential 0,17910620335162064
Logic Temporal 0,11891767800211263
Logic Upward Monotone 0,7009124021507408
Logic Downward Monotone 0,16146816171752817
Logic Non-Monotonic 0,39405520311955033
Knowledge Common Sense 0,3600431599767098
Knowledge World Knowledge 0,4447466812418805

Производительность:

Датасет Speed RAM
LiDiRus - -
RCB - -
PARus - -
MuSeRC - -
TERRa - -
RUSSE - -
RWSD - -
DaNetQA - -
RuCoS - -