Overview
T5-Base v1.1 model trained to generate hypotheses given a premise and a label. Below the settings used to train it.
Experiment configurations
├── datasets
│ └── snli_train:
│ dataset_name: snli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names: null
│ val_subset_names: validation
│ test_subset_names: none
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│ anli_train:
│ dataset_name: anli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names:
│ - train_r1
│ - train_r2
│ - train_r3
│ val_subset_names:
│ - dev_r1
│ - dev_r2
│ - dev_r3
│ test_subset_names: none
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│ mnli_train:
│ dataset_name: multi_nli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names: null
│ val_subset_names: validation_matched
│ test_subset_names: none
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│ snli:
│ dataset_name: snli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names: none
│ val_subset_names: none
│ test_subset_names: null
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│ anli:
│ dataset_name: anli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names: none
│ val_subset_names: none
│ test_subset_names:
│ - test_r1
│ - test_r2
│ - test_r3
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│ mnli:
│ dataset_name: multi_nli
│ dataset_config_name: null
│ cache_dir: null
│ input_fields:
│ - premise
│ - hypothesis
│ target_field: label
│ train_subset_names: none
│ val_subset_names: none
│ test_subset_names: validation_mismatched
│ train_val_split: null
│ limit_train_samples: null
│ limit_val_samples: null
│ limit_test_samples: null
│ sampling_kwargs:
│ sampling_strategy: random
│ seed: 42
│ replace: false
│ align_labels_with_mapping: null
│ avoid_consistency_check: false
│ predict_label_mapping: null
│
├── data
│ └── _target_: src.task.nli.data.NLIGenerationData.from_config
│ main_dataset_name: null
│ use_additional_as_test: null
│ dataloader:
│ batch_size: 96
│ eval_batch_size: 96
│ num_workers: 8
│ pin_memory: true
│ drop_last: false
│ persistent_workers: false
│ shuffle: true
│ seed_dataloader: 42
│ replacement: false
│ processing:
│ preprocessing_num_workers: 8
│ preprocessing_batch_size: 1000
│ load_from_cache_file: true
│ padding: longest
│ truncation: longest_first
│ max_source_length: 128
│ max_target_length: 128
│ template: 'premise: $premise $label hypothesis: '
│ tokenizer:
│ _target_: transformers.AutoTokenizer.from_pretrained
│ pretrained_model_name_or_path: pietrolesci/t5-v1_1-base_nli_gen
│ use_fast: true
│
├── task
│ └── optimizer:
│ name: Adafactor
│ lr: 0.001
│ weight_decay: 0.0
│ no_decay:
│ - bias
│ - LayerNorm.weight
│ decay_rate: -0.8
│ clip_threshold: 1.0
│ relative_step: false
│ scale_parameter: false
│ warmup_init: false
│ scheduler:
│ name: constant_schedule
│ model:
│ model_name_or_path: pietrolesci/t5-v1_1-base_nli_gen
│ checkpoint_path: null
│ freeze: false
│ seed_init_weight: 42
│ _target_: src.task.nli.NLIGenerationTask.from_config
│ generation:
│ generation_max_length: 128
│ generation_min_length: 3
│ do_sample: true
│ early_stopping: false
│ num_beams: 1
│ temperature: 1.0
│ top_k: 50
│ top_p: 0.95
│ repetition_penalty: null
│ length_penalty: null
│ no_repeat_ngram_size: null
│ encoder_no_repeat_ngram_size: null
│ num_return_sequences: 1
│ max_time: null
│ max_new_tokens: null
│ decoder_start_token_id: null
│ use_cache: null
│ num_beam_groups: null
│ diversity_penalty: null
│
├── trainer
│ └── _target_: pytorch_lightning.Trainer
│ callbacks:
│ lr_monitor:
│ _target_: pytorch_lightning.callbacks.LearningRateMonitor
│ logging_interval: step
│ log_momentum: false
│ model_checkpoint:
│ _target_: pytorch_lightning.callbacks.ModelCheckpoint
│ dirpath: ./checkpoints/
│ filename: nli_generator_sma-epoch={epoch:02d}-val_loss={val/aggregat
│ monitor: val/aggregated_loss
│ mode: min
│ verbose: false
│ save_last: true
│ save_top_k: 1
│ auto_insert_metric_name: false
│ save_on_train_epoch_end: false
│ rich_model_summary:
│ _target_: pytorch_lightning.callbacks.RichModelSummary
│ max_depth: 1
│ log_grad_norm:
│ _target_: src.core.callbacks.LogGradNorm
│ norm_type: 2
│ group_separator: /
│ only_total: true
│ on_step: true
│ on_epoch: false
│ prog_bar: true
│ log_generated_text:
│ _target_: src.core.callbacks.GenerateAndLogText
│ dirpath: ./generated_text
│ type: generated_text
│ pop_keys_after_logging: true
│ on_train: false
│ on_validation: false
│ on_test: true
│ log_to_wandb: true
│ wandb_log_dataset_sizes:
│ _target_: src.core.callbacks.WandbLogDatasetSizes
│ logger:
│ wandb:
│ _target_: pytorch_lightning.loggers.WandbLogger
│ project: nli_debiasing
│ entity: team_brushino
│ name: nli_generator_sma
│ save_dir: ./
│ offline: false
│ log_model: false
│ group: generator
│ job_type: genearator_training
│ tags:
│ - nli_generator_sma
│ - seed=42
│ - seed_dataloader=42
│ notes: nli_generator_sma_time=01-37-04
│ enable_checkpointing: true
│ enable_progress_bar: true
│ enable_model_summary: true
│ gradient_clip_val: 6
│ gradient_clip_algorithm: null
│ accelerator: gpu
│ devices: auto
│ gpus: null
│ auto_select_gpus: true
│ accumulate_grad_batches: 1
│ max_epochs: 2
│ min_epochs: 1
│ max_steps: -1
│ min_steps: null
│ max_time: null
│ num_sanity_val_steps: 2
│ overfit_batches: 0.0
│ fast_dev_run: false
│ limit_train_batches: 1.0
│ limit_val_batches: 1.0
│ limit_test_batches: 1.0
│ profiler: null
│ detect_anomaly: false
│ deterministic: false
│ check_val_every_n_epoch: 1
│ val_check_interval: 0.5
│ log_every_n_steps: 1
│ move_metrics_to_cpu: false
│
└── training
└── run_val_before_fit: false
run_val_after_fit: false
run_test_before_fit: false
run_test_after_fit: true
lr: 0.001
seed: 42
show_batch: false
batch_size: 96
eval_batch_size: 96
num_workers: 8
pin_memory: true
drop_last: false
persistent_workers: false
shuffle: true
seed_dataloader: 42
ignore_warnings: true
experiment_name: nli_generator_sma
