README.md
| 1 | ## ELECTRA_large_discriminator language model fine-tuned on SQuAD2.0 |
| 2 | |
| 3 | ### with the following results: |
| 4 | |
| 5 | ``` |
| 6 | "exact": 87.09677419354838, |
| 7 | "f1": 89.98343832723452, |
| 8 | "total": 11873, |
| 9 | "HasAns_exact": 84.66599190283401, |
| 10 | "HasAns_f1": 90.44759839056285, |
| 11 | "HasAns_total": 5928, |
| 12 | "NoAns_exact": 89.52060555088309, |
| 13 | "NoAns_f1": 89.52060555088309, |
| 14 | "NoAns_total": 5945, |
| 15 | "best_exact": 87.09677419354838, |
| 16 | "best_exact_thresh": 0.0, |
| 17 | "best_f1": 89.98343832723432, |
| 18 | "best_f1_thresh": 0.0 |
| 19 | ``` |
| 20 | ### from script: |
| 21 | ``` |
| 22 | python ${EXAMPLES}/run_squad.py \ |
| 23 | --model_type electra \ |
| 24 | --model_name_or_path google/electra-large-discriminator \ |
| 25 | --do_train \ |
| 26 | --do_eval \ |
| 27 | --train_file ${SQUAD}/train-v2.0.json \ |
| 28 | --predict_file ${SQUAD}/dev-v2.0.json \ |
| 29 | --version_2_with_negative \ |
| 30 | --do_lower_case \ |
| 31 | --num_train_epochs 3 \ |
| 32 | --warmup_steps 306 \ |
| 33 | --weight_decay 0.01 \ |
| 34 | --learning_rate 3e-5 \ |
| 35 | --max_grad_norm 0.5 \ |
| 36 | --adam_epsilon 1e-6 \ |
| 37 | --max_seq_length 512 \ |
| 38 | --doc_stride 128 \ |
| 39 | --per_gpu_train_batch_size 8 \ |
| 40 | --gradient_accumulation_steps 16 \ |
| 41 | --per_gpu_eval_batch_size 128 \ |
| 42 | --fp16 \ |
| 43 | --fp16_opt_level O1 \ |
| 44 | --threads 12 \ |
| 45 | --logging_steps 50 \ |
| 46 | --save_steps 1000 \ |
| 47 | --overwrite_output_dir \ |
| 48 | --output_dir ${MODEL_PATH} |
| 49 | ``` |
| 50 | ### using the following system & software: |
| 51 | ``` |
| 52 | Transformers: 2.11.0 |
| 53 | PyTorch: 1.5.0 |
| 54 | TensorFlow: 2.2.0 |
| 55 | Python: 3.8.1 |
| 56 | OS/Platform: Linux-5.3.0-59-generic-x86_64-with-glibc2.10 |
| 57 | CPU/GPU: Intel i9-9900K / NVIDIA Titan RTX 24GB |
| 58 | ``` |
| 59 | |