Initial commit

README.md

@@ -1,3 +1,134 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- kresnik/zeroth_korean
+- generated_from_trainer
+datasets:
+- zeroth_korean_asr
+model-index:
+- name: ''
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# 
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the KRESNIK/ZEROTH_KOREAN - CLEAN dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.2089
+- Wer: 0.2954
+- Cer: 0.0953
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 7.5e-05
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 32
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 2000
+- num_epochs: 50.0
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    | Cer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|:------:|
+| 19.7138       | 0.72  | 500   | 19.6427         | 1.0    | 1.0    |
+| 4.8039        | 1.44  | 1000  | 4.7842          | 1.0    | 1.0    |
+| 4.5619        | 2.16  | 1500  | 4.5608          | 0.9992 | 0.9598 |
+| 4.254         | 2.88  | 2000  | 4.2729          | 0.9955 | 0.9063 |
+| 4.1905        | 3.6   | 2500  | 4.2257          | 0.9903 | 0.8758 |
+| 4.0683        | 4.32  | 3000  | 3.9294          | 0.9937 | 0.7911 |
+| 3.486         | 5.04  | 3500  | 2.7045          | 1.0012 | 0.5934 |
+| 2.946         | 5.75  | 4000  | 1.9691          | 0.9425 | 0.4634 |
+| 2.634         | 6.47  | 4500  | 1.5212          | 0.8807 | 0.3850 |
+| 2.4066        | 7.19  | 5000  | 1.2551          | 0.8177 | 0.3601 |
+| 2.2651        | 7.91  | 5500  | 1.0423          | 0.7650 | 0.3039 |
+| 2.1828        | 8.63  | 6000  | 0.9599          | 0.7273 | 0.3106 |
+| 2.1023        | 9.35  | 6500  | 0.9482          | 0.7161 | 0.3063 |
+| 2.0536        | 10.07 | 7000  | 0.8242          | 0.6767 | 0.2860 |
+| 1.9803        | 10.79 | 7500  | 0.7643          | 0.6563 | 0.2637 |
+| 1.9468        | 11.51 | 8000  | 0.7319          | 0.6441 | 0.2505 |
+| 1.9178        | 12.23 | 8500  | 0.6937          | 0.6320 | 0.2489 |
+| 1.8515        | 12.95 | 9000  | 0.6443          | 0.6053 | 0.2196 |
+| 1.8083        | 13.67 | 9500  | 0.6286          | 0.6122 | 0.2148 |
+| 1.819         | 14.39 | 10000 | 0.6015          | 0.5986 | 0.2074 |
+| 1.7684        | 15.11 | 10500 | 0.5682          | 0.5741 | 0.1982 |
+| 1.7195        | 15.83 | 11000 | 0.5385          | 0.5592 | 0.2007 |
+| 1.7044        | 16.55 | 11500 | 0.5362          | 0.5524 | 0.2097 |
+| 1.6879        | 17.27 | 12000 | 0.5119          | 0.5489 | 0.2083 |
+| 1.656         | 17.98 | 12500 | 0.4990          | 0.5362 | 0.1968 |
+| 1.6122        | 18.7  | 13000 | 0.4561          | 0.5092 | 0.1900 |
+| 1.5919        | 19.42 | 13500 | 0.4778          | 0.5225 | 0.1975 |
+| 1.5896        | 20.14 | 14000 | 0.4563          | 0.5098 | 0.1859 |
+| 1.5589        | 20.86 | 14500 | 0.4362          | 0.4940 | 0.1725 |
+| 1.5353        | 21.58 | 15000 | 0.4140          | 0.4826 | 0.1580 |
+| 1.5441        | 22.3  | 15500 | 0.4031          | 0.4742 | 0.1550 |
+| 1.5116        | 23.02 | 16000 | 0.3916          | 0.4748 | 0.1545 |
+| 1.4731        | 23.74 | 16500 | 0.3841          | 0.4810 | 0.1542 |
+| 1.4647        | 24.46 | 17000 | 0.3752          | 0.4524 | 0.1475 |
+| 1.4328        | 25.18 | 17500 | 0.3587          | 0.4476 | 0.1461 |
+| 1.4129        | 25.9  | 18000 | 0.3429          | 0.4242 | 0.1366 |
+| 1.4062        | 26.62 | 18500 | 0.3450          | 0.4251 | 0.1355 |
+| 1.3928        | 27.34 | 19000 | 0.3297          | 0.4145 | 0.1322 |
+| 1.3906        | 28.06 | 19500 | 0.3210          | 0.4185 | 0.1336 |
+| 1.358         | 28.78 | 20000 | 0.3131          | 0.3970 | 0.1275 |
+| 1.3445        | 29.5  | 20500 | 0.3069          | 0.3920 | 0.1276 |
+| 1.3159        | 30.22 | 21000 | 0.3035          | 0.3961 | 0.1255 |
+| 1.3044        | 30.93 | 21500 | 0.2952          | 0.3854 | 0.1242 |
+| 1.3034        | 31.65 | 22000 | 0.2966          | 0.3772 | 0.1227 |
+| 1.2963        | 32.37 | 22500 | 0.2844          | 0.3706 | 0.1208 |
+| 1.2765        | 33.09 | 23000 | 0.2841          | 0.3567 | 0.1173 |
+| 1.2438        | 33.81 | 23500 | 0.2734          | 0.3552 | 0.1137 |
+| 1.2487        | 34.53 | 24000 | 0.2703          | 0.3502 | 0.1118 |
+| 1.2249        | 35.25 | 24500 | 0.2650          | 0.3484 | 0.1142 |
+| 1.2229        | 35.97 | 25000 | 0.2584          | 0.3374 | 0.1097 |
+| 1.2374        | 36.69 | 25500 | 0.2568          | 0.3337 | 0.1095 |
+| 1.2153        | 37.41 | 26000 | 0.2494          | 0.3327 | 0.1071 |
+| 1.1925        | 38.13 | 26500 | 0.2518          | 0.3366 | 0.1077 |
+| 1.1908        | 38.85 | 27000 | 0.2437          | 0.3272 | 0.1057 |
+| 1.1858        | 39.57 | 27500 | 0.2396          | 0.3265 | 0.1044 |
+| 1.1808        | 40.29 | 28000 | 0.2373          | 0.3156 | 0.1028 |
+| 1.1842        | 41.01 | 28500 | 0.2356          | 0.3152 | 0.1026 |
+| 1.1668        | 41.73 | 29000 | 0.2319          | 0.3188 | 0.1025 |
+| 1.1448        | 42.45 | 29500 | 0.2293          | 0.3099 | 0.0995 |
+| 1.1327        | 43.17 | 30000 | 0.2265          | 0.3047 | 0.0979 |
+| 1.1307        | 43.88 | 30500 | 0.2222          | 0.3078 | 0.0989 |
+| 1.1419        | 44.6  | 31000 | 0.2215          | 0.3038 | 0.0981 |
+| 1.1231        | 45.32 | 31500 | 0.2193          | 0.3013 | 0.0972 |
+| 1.139         | 46.04 | 32000 | 0.2162          | 0.3007 | 0.0968 |
+| 1.1114        | 46.76 | 32500 | 0.2122          | 0.2982 | 0.0960 |
+| 1.111         | 47.48 | 33000 | 0.2125          | 0.2946 | 0.0948 |
+| 1.0982        | 48.2  | 33500 | 0.2099          | 0.2957 | 0.0953 |
+| 1.109         | 48.92 | 34000 | 0.2092          | 0.2955 | 0.0955 |
+| 1.0905        | 49.64 | 34500 | 0.2088          | 0.2954 | 0.0953 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.10.3

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 1205, "</s>": 1206}

all_results.json

@@ -0,0 +1,15 @@
+{
+    "epoch": 50.0,
+    "eval_cer": 0.09533931664304834,
+    "eval_loss": 0.20887407660484314,
+    "eval_runtime": 41.7123,
+    "eval_samples": 456,
+    "eval_samples_per_second": 10.932,
+    "eval_steps_per_second": 1.367,
+    "eval_wer": 0.2953790395650861,
+    "train_loss": 2.2316733406121783,
+    "train_runtime": 114311.9751,
+    "train_samples": 22262,
+    "train_samples_per_second": 9.737,
+    "train_steps_per_second": 0.304
+}

config.json

@@ -0,0 +1,107 @@
+{
+  "_name_or_path": "facebook/wav2vec2-xls-r-300m",
+  "activation_dropout": 0.1,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 768,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": false,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.0,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.0,
+  "mask_feature_length": 64,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.25,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.75,
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "num_negatives": 100,
+  "output_hidden_size": 1024,
+  "pad_token_id": 1204,
+  "proj_codevector_dim": 768,
+  "tdnn_dilation": [
+    1,
+    2,
+    3,
+    1,
+    1
+  ],
+  "tdnn_dim": [
+    512,
+    512,
+    512,
+    512,
+    1500
+  ],
+  "tdnn_kernel": [
+    5,
+    3,
+    3,
+    1,
+    1
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.17.0.dev0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 1207,
+  "xvector_output_dim": 512
+}

eval_results.json

@@ -0,0 +1,10 @@
+{
+    "epoch": 50.0,
+    "eval_cer": 0.09533931664304834,
+    "eval_loss": 0.20887407660484314,
+    "eval_runtime": 41.7123,
+    "eval_samples": 456,
+    "eval_samples_per_second": 10.932,
+    "eval_steps_per_second": 1.367,
+    "eval_wer": 0.2953790395650861
+}

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bc8f067dea3b8b2f7be8445f2df742139fa6fedc42dc96052e1f03903ed3434
+size 1266872433

run.sh

@@ -0,0 +1,34 @@
+python run_speech_recognition_ctc.py \
+	--dataset_name="kresnik/zeroth_korean" \
+	--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
+	--dataset_config_name="clean" \
+	--output_dir="./" \
+	--overwrite_output_dir \
+	--num_train_epochs="50" \
+	--per_device_train_batch_size="8" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="4" \
+	--learning_rate="7.5e-5" \
+	--warmup_steps="2000" \
+	--length_column_name="input_length" \
+	--evaluation_strategy="steps" \
+	--text_column_name="text" \
+	--chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
+	--save_steps="500" \
+	--eval_steps="500" \
+	--logging_steps="100" \
+	--layerdrop="0.0" \
+	--activation_dropout="0.1" \
+	--save_total_limit="3" \
+	--freeze_feature_encoder \
+	--feat_proj_dropout="0.0" \
+	--mask_time_prob="0.75" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.25" \
+	--mask_feature_length="64" \
+	--gradient_checkpointing \
+	--use_auth_token \
+	--fp16 \
+	--group_by_length \
+	--do_train --do_eval \
+	--push_to_hub

run_speech_recognition_ctc.py

@@ -0,0 +1,829 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.17.0.dev0")
+
+require_version(
+    "datasets>=1.13.3",
+    "To fix: pip install -r examples/pytorch/text-classification/requirements.txt",
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={
+            "help": "Path to pretrained model or model identifier from huggingface.co/models"
+        }
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"
+        },
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Where do you want to store the pretrained models downloaded from huggingface.co"
+        },
+    )
+    freeze_feature_encoder: bool = field(
+        default=True,
+        metadata={"help": "Whether to freeze the feature encoder layers of the model."},
+    )
+    attention_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout ratio for the attention probabilities."},
+    )
+    activation_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout ratio for activations inside the fully connected layer."
+        },
+    )
+    feat_proj_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the projected features."}
+    )
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+            "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(
+        default=0.0, metadata={"help": "The LayerDrop probability."}
+    )
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean",
+        metadata={
+            "help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."
+        },
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: str = field(
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."
+        }
+    )
+    dataset_config_name: str = field(
+        default=None,
+        metadata={
+            "help": "The configuration name of the dataset to use (via the datasets library)."
+        },
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'test'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={
+            "help": "The name of the dataset column containing the audio data. Defaults to 'audio'"
+        },
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={
+            "help": "The name of the dataset column containing the text data. Defaults to 'text'"
+        },
+    )
+    overwrite_cache: bool = field(
+        default=False,
+        metadata={"help": "Overwrite the cached preprocessed datasets or not."},
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer", "cer"],
+        metadata={
+            "help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"
+        },
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0,
+        metadata={
+            "help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"
+        },
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]", metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]", metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|", metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(
+        self, features: List[Dict[str, Union[List[int], torch.Tensor]]]
+    ) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [
+            {"input_values": feature["input_values"]} for feature in features
+        ]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(
+            labels_batch.attention_mask.ne(1), -100
+        )
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]),
+        vocabs.values(),
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser(
+        (ModelArguments, DataTrainingArguments, TrainingArguments)
+    )
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(
+            json_file=os.path.abspath(sys.argv[1])
+        )
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if (
+        os.path.isdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(
+        logging.INFO if is_main_process(training_args.local_rank) else logging.WARN
+    )
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(
+                range(data_args.max_train_samples)
+            )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(
+                range(data_args.max_eval_samples)
+            )
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    chars_to_ignore_regex = (
+        f'[{"".join(data_args.chars_to_ignore)}]'
+        if data_args.chars_to_ignore is not None
+        else None
+    )
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = (
+                re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+            )
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(
+        desc="dataset map special characters removal"
+    ):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type
+            if config.tokenizer_class is None
+            else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = (
+        next(iter(raw_datasets.values()))
+        .features[data_args.audio_column_name]
+        .sampling_rate
+    )
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name,
+            datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate),
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = (
+        data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    )
+    min_input_length = (
+        data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    )
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(
+            sample["array"], sampling_rate=sample["sampling_rate"]
+        )
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(
+            f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
+        )
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {
+            k: v.compute(predictions=pred_str, references=label_str)
+            for k, v in eval_metrics.items()
+        }
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(
+            max_train_samples, len(vectorized_datasets["train"])
+        )
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples
+            if data_args.max_eval_samples is not None
+            else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(
+            max_eval_samples, len(vectorized_datasets["eval"])
+        )
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = (
+        data_args.dataset_config_name
+        if data_args.dataset_config_name is not None
+        else "na"
+    )
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

runs/Jan31_07-15-59_job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6/1643613501.488685/events.out.tfevents.1643613501.job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6.1151936.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28a6ab5fcdee80fd31c69dc157696d3063a1cc27099a2452f6695c51cba48628
+size 4753

runs/Jan31_07-15-59_job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6/events.out.tfevents.1643613501.job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6.1151936.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e586cc74e2f42c18b87325dcb4cc447cd7318cc76addfa59195430faa1ea2ea5
+size 85518

runs/Jan31_07-15-59_job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6/events.out.tfevents.1643727998.job-2c68f48a-2d5d-4013-9043-3f2cb25f3ff6.1151936.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fcf7744c50287e982c2028e14e51bc27110121c60f9a7d4ec7fc26a06f86232
+size 412

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 50.0,
+    "train_loss": 2.2316733406121783,
+    "train_runtime": 114311.9751,
+    "train_samples": 22262,
+    "train_samples_per_second": 9.737,
+    "train_steps_per_second": 0.304
+}

trainer_state.json

@@ -0,0 +1,2797 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 49.99892202659001,
+  "global_step": 34750,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.14,
+      "learning_rate": 3.675e-06,
+      "loss": 47.2908,
+      "step": 100
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 7.425e-06,
+      "loss": 33.9125,
+      "step": 200
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1.1174999999999999e-05,
+      "loss": 26.6068,
+      "step": 300
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1.4925e-05,
+      "loss": 23.2775,
+      "step": 400
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1.8675e-05,
+      "loss": 19.7138,
+      "step": 500
+    },
+    {
+      "epoch": 0.72,
+      "eval_cer": 1.0,
+      "eval_loss": 19.642736434936523,
+      "eval_runtime": 41.3907,
+      "eval_samples_per_second": 11.017,
+      "eval_steps_per_second": 1.377,
+      "eval_wer": 1.0,
+      "step": 500
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 2.2424999999999996e-05,
+      "loss": 15.7715,
+      "step": 600
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 2.6174999999999996e-05,
+      "loss": 11.4061,
+      "step": 700
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 2.9925e-05,
+      "loss": 7.4329,
+      "step": 800
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 3.3675e-05,
+      "loss": 5.3081,
+      "step": 900
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 3.7424999999999995e-05,
+      "loss": 4.8039,
+      "step": 1000
+    },
+    {
+      "epoch": 1.44,
+      "eval_cer": 1.0,
+      "eval_loss": 4.784187316894531,
+      "eval_runtime": 42.2256,
+      "eval_samples_per_second": 10.799,
+      "eval_steps_per_second": 1.35,
+      "eval_wer": 1.0,
+      "step": 1000
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 4.1175e-05,
+      "loss": 4.762,
+      "step": 1100
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 4.4924999999999994e-05,
+      "loss": 4.6928,
+      "step": 1200
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 4.8675e-05,
+      "loss": 4.6292,
+      "step": 1300
+    },
+    {
+      "epoch": 2.01,
+      "learning_rate": 5.2424999999999994e-05,
+      "loss": 4.6321,
+      "step": 1400
+    },
+    {
+      "epoch": 2.16,
+      "learning_rate": 5.6175e-05,
+      "loss": 4.5619,
+      "step": 1500
+    },
+    {
+      "epoch": 2.16,
+      "eval_cer": 0.9598094788222327,
+      "eval_loss": 4.560794830322266,
+      "eval_runtime": 41.0352,
+      "eval_samples_per_second": 11.112,
+      "eval_steps_per_second": 1.389,
+      "eval_wer": 0.9992449411054062,
+      "step": 1500
+    },
+    {
+      "epoch": 2.3,
+      "learning_rate": 5.9925e-05,
+      "loss": 4.4704,
+      "step": 1600
+    },
+    {
+      "epoch": 2.45,
+      "learning_rate": 6.367499999999999e-05,
+      "loss": 4.3806,
+      "step": 1700
+    },
+    {
+      "epoch": 2.59,
+      "learning_rate": 6.7425e-05,
+      "loss": 4.3092,
+      "step": 1800
+    },
+    {
+      "epoch": 2.73,
+      "learning_rate": 7.1175e-05,
+      "loss": 4.2794,
+      "step": 1900
+    },
+    {
+      "epoch": 2.88,
+      "learning_rate": 7.492499999999999e-05,
+      "loss": 4.254,
+      "step": 2000
+    },
+    {
+      "epoch": 2.88,
+      "eval_cer": 0.906274602424815,
+      "eval_loss": 4.272861003875732,
+      "eval_runtime": 40.8387,
+      "eval_samples_per_second": 11.166,
+      "eval_steps_per_second": 1.396,
+      "eval_wer": 0.9954696466324373,
+      "step": 2000
+    },
+    {
+      "epoch": 3.02,
+      "learning_rate": 7.477557251908395e-05,
+      "loss": 4.2616,
+      "step": 2100
+    },
+    {
+      "epoch": 3.17,
+      "learning_rate": 7.454656488549618e-05,
+      "loss": 4.2184,
+      "step": 2200
+    },
+    {
+      "epoch": 3.31,
+      "learning_rate": 7.43175572519084e-05,
+      "loss": 4.227,
+      "step": 2300
+    },
+    {
+      "epoch": 3.45,
+      "learning_rate": 7.408854961832061e-05,
+      "loss": 4.1985,
+      "step": 2400
+    },
+    {
+      "epoch": 3.6,
+      "learning_rate": 7.385954198473281e-05,
+      "loss": 4.1905,
+      "step": 2500
+    },
+    {
+      "epoch": 3.6,
+      "eval_cer": 0.8758463234136357,
+      "eval_loss": 4.225706100463867,
+      "eval_runtime": 40.6017,
+      "eval_samples_per_second": 11.231,
+      "eval_steps_per_second": 1.404,
+      "eval_wer": 0.9903352461491997,
+      "step": 2500
+    },
+    {
+      "epoch": 3.74,
+      "learning_rate": 7.36328244274809e-05,
+      "loss": 4.1873,
+      "step": 2600
+    },
+    {
+      "epoch": 3.88,
+      "learning_rate": 7.340381679389312e-05,
+      "loss": 4.1615,
+      "step": 2700
+    },
+    {
+      "epoch": 4.03,
+      "learning_rate": 7.317480916030534e-05,
+      "loss": 4.157,
+      "step": 2800
+    },
+    {
+      "epoch": 4.17,
+      "learning_rate": 7.294580152671756e-05,
+      "loss": 4.1124,
+      "step": 2900
+    },
+    {
+      "epoch": 4.32,
+      "learning_rate": 7.271679389312976e-05,
+      "loss": 4.0683,
+      "step": 3000
+    },
+    {
+      "epoch": 4.32,
+      "eval_cer": 0.7911352542906629,
+      "eval_loss": 3.929443120956421,
+      "eval_runtime": 41.2715,
+      "eval_samples_per_second": 11.049,
+      "eval_steps_per_second": 1.381,
+      "eval_wer": 0.9936575052854123,
+      "step": 3000
+    },
+    {
+      "epoch": 4.46,
+      "learning_rate": 7.248778625954197e-05,
+      "loss": 4.0704,
+      "step": 3100
+    },
+    {
+      "epoch": 4.6,
+      "learning_rate": 7.22587786259542e-05,
+      "loss": 3.9616,
+      "step": 3200
+    },
+    {
+      "epoch": 4.75,
+      "learning_rate": 7.202977099236641e-05,
+      "loss": 3.7798,
+      "step": 3300
+    },
+    {
+      "epoch": 4.89,
+      "learning_rate": 7.180076335877862e-05,
+      "loss": 3.6123,
+      "step": 3400
+    },
+    {
+      "epoch": 5.04,
+      "learning_rate": 7.15740458015267e-05,
+      "loss": 3.486,
+      "step": 3500
+    },
+    {
+      "epoch": 5.04,
+      "eval_cer": 0.5933711226578492,
+      "eval_loss": 2.704545497894287,
+      "eval_runtime": 40.8408,
+      "eval_samples_per_second": 11.165,
+      "eval_steps_per_second": 1.396,
+      "eval_wer": 1.0012080942313502,
+      "step": 3500
+    },
+    {
+      "epoch": 5.18,
+      "learning_rate": 7.134503816793892e-05,
+      "loss": 3.3283,
+      "step": 3600
+    },
+    {
+      "epoch": 5.32,
+      "learning_rate": 7.111603053435114e-05,
+      "loss": 3.2091,
+      "step": 3700
+    },
+    {
+      "epoch": 5.47,
+      "learning_rate": 7.088931297709923e-05,
+      "loss": 3.1158,
+      "step": 3800
+    },
+    {
+      "epoch": 5.61,
+      "learning_rate": 7.066030534351145e-05,
+      "loss": 2.9983,
+      "step": 3900
+    },
+    {
+      "epoch": 5.75,
+      "learning_rate": 7.043129770992365e-05,
+      "loss": 2.946,
+      "step": 4000
+    },
+    {
+      "epoch": 5.75,
+      "eval_cer": 0.4634309557549992,
+      "eval_loss": 1.9690674543380737,
+      "eval_runtime": 40.879,
+      "eval_samples_per_second": 11.155,
+      "eval_steps_per_second": 1.394,
+      "eval_wer": 0.942464512231954,
+      "step": 4000
+    },
+    {
+      "epoch": 5.9,
+      "learning_rate": 7.020229007633587e-05,
+      "loss": 2.8545,
+      "step": 4100
+    },
+    {
+      "epoch": 6.04,
+      "learning_rate": 6.997328244274808e-05,
+      "loss": 2.8092,
+      "step": 4200
+    },
+    {
+      "epoch": 6.19,
+      "learning_rate": 6.97442748091603e-05,
+      "loss": 2.7229,
+      "step": 4300
+    },
+    {
+      "epoch": 6.33,
+      "learning_rate": 6.951526717557252e-05,
+      "loss": 2.7053,
+      "step": 4400
+    },
+    {
+      "epoch": 6.47,
+      "learning_rate": 6.928625954198472e-05,
+      "loss": 2.634,
+      "step": 4500
+    },
+    {
+      "epoch": 6.47,
+      "eval_cer": 0.38501810738466385,
+      "eval_loss": 1.521231770515442,
+      "eval_runtime": 41.5435,
+      "eval_samples_per_second": 10.976,
+      "eval_steps_per_second": 1.372,
+      "eval_wer": 0.880700694654183,
+      "step": 4500
+    },
+    {
+      "epoch": 6.62,
+      "learning_rate": 6.905725190839693e-05,
+      "loss": 2.5996,
+      "step": 4600
+    },
+    {
+      "epoch": 6.76,
+      "learning_rate": 6.882824427480916e-05,
+      "loss": 2.5472,
+      "step": 4700
+    },
+    {
+      "epoch": 6.91,
+      "learning_rate": 6.859923664122137e-05,
+      "loss": 2.4959,
+      "step": 4800
+    },
+    {
+      "epoch": 7.05,
+      "learning_rate": 6.837022900763359e-05,
+      "loss": 2.4554,
+      "step": 4900
+    },
+    {
+      "epoch": 7.19,
+      "learning_rate": 6.814122137404579e-05,
+      "loss": 2.4066,
+      "step": 5000
+    },
+    {
+      "epoch": 7.19,
+      "eval_cer": 0.36014013541174617,
+      "eval_loss": 1.2550952434539795,
+      "eval_runtime": 41.0408,
+      "eval_samples_per_second": 11.111,
+      "eval_steps_per_second": 1.389,
+      "eval_wer": 0.8177287828450619,
+      "step": 5000
+    },
+    {
+      "epoch": 7.34,
+      "learning_rate": 6.791221374045801e-05,
+      "loss": 2.3768,
+      "step": 5100
+    },
+    {
+      "epoch": 7.48,
+      "learning_rate": 6.768320610687023e-05,
+      "loss": 2.3557,
+      "step": 5200
+    },
+    {
+      "epoch": 7.63,
+      "learning_rate": 6.745419847328244e-05,
+      "loss": 2.3109,
+      "step": 5300
+    },
+    {
+      "epoch": 7.77,
+      "learning_rate": 6.722519083969465e-05,
+      "loss": 2.2953,
+      "step": 5400
+    },
+    {
+      "epoch": 7.91,
+      "learning_rate": 6.699618320610687e-05,
+      "loss": 2.2651,
+      "step": 5500
+    },
+    {
+      "epoch": 7.91,
+      "eval_cer": 0.30392851519445757,
+      "eval_loss": 1.0423332452774048,
+      "eval_runtime": 40.9098,
+      "eval_samples_per_second": 11.146,
+      "eval_steps_per_second": 1.393,
+      "eval_wer": 0.7650256720024162,
+      "step": 5500
+    },
+    {
+      "epoch": 8.06,
+      "learning_rate": 6.676717557251908e-05,
+      "loss": 2.2589,
+      "step": 5600
+    },
+    {
+      "epoch": 8.2,
+      "learning_rate": 6.654045801526718e-05,
+      "loss": 2.2122,
+      "step": 5700
+    },
+    {
+      "epoch": 8.34,
+      "learning_rate": 6.631145038167939e-05,
+      "loss": 2.2017,
+      "step": 5800
+    },
+    {
+      "epoch": 8.49,
+      "learning_rate": 6.60824427480916e-05,
+      "loss": 2.1814,
+      "step": 5900
+    },
+    {
+      "epoch": 8.63,
+      "learning_rate": 6.58534351145038e-05,
+      "loss": 2.1828,
+      "step": 6000
+    },
+    {
+      "epoch": 8.63,
+      "eval_cer": 0.3106203747441348,
+      "eval_loss": 0.9598844051361084,
+      "eval_runtime": 41.4743,
+      "eval_samples_per_second": 10.995,
+      "eval_steps_per_second": 1.374,
+      "eval_wer": 0.7272727272727273,
+      "step": 6000
+    },
+    {
+      "epoch": 8.78,
+      "learning_rate": 6.562442748091603e-05,
+      "loss": 2.1714,
+      "step": 6100
+    },
+    {
+      "epoch": 8.92,
+      "learning_rate": 6.539541984732824e-05,
+      "loss": 2.1422,
+      "step": 6200
+    },
+    {
+      "epoch": 9.06,
+      "learning_rate": 6.516641221374046e-05,
+      "loss": 2.1546,
+      "step": 6300
+    },
+    {
+      "epoch": 9.21,
+      "learning_rate": 6.493740458015267e-05,
+      "loss": 2.12,
+      "step": 6400
+    },
+    {
+      "epoch": 9.35,
+      "learning_rate": 6.470839694656488e-05,
+      "loss": 2.1023,
+      "step": 6500
+    },
+    {
+      "epoch": 9.35,
+      "eval_cer": 0.30632971185640057,
+      "eval_loss": 0.9481843113899231,
+      "eval_runtime": 41.1867,
+      "eval_samples_per_second": 11.072,
+      "eval_steps_per_second": 1.384,
+      "eval_wer": 0.7160978556327393,
+      "step": 6500
+    },
+    {
+      "epoch": 9.5,
+      "learning_rate": 6.44793893129771e-05,
+      "loss": 2.1104,
+      "step": 6600
+    },
+    {
+      "epoch": 9.64,
+      "learning_rate": 6.425038167938931e-05,
+      "loss": 2.0879,
+      "step": 6700
+    },
+    {
+      "epoch": 9.78,
+      "learning_rate": 6.402137404580152e-05,
+      "loss": 2.0724,
+      "step": 6800
+    },
+    {
+      "epoch": 9.93,
+      "learning_rate": 6.379236641221374e-05,
+      "loss": 2.0622,
+      "step": 6900
+    },
+    {
+      "epoch": 10.07,
+      "learning_rate": 6.356335877862595e-05,
+      "loss": 2.0536,
+      "step": 7000
+    },
+    {
+      "epoch": 10.07,
+      "eval_cer": 0.28597858604944104,
+      "eval_loss": 0.8241907954216003,
+      "eval_runtime": 41.2837,
+      "eval_samples_per_second": 11.046,
+      "eval_steps_per_second": 1.381,
+      "eval_wer": 0.6766837813349441,
+      "step": 7000
+    },
+    {
+      "epoch": 10.22,
+      "learning_rate": 6.333435114503816e-05,
+      "loss": 2.0258,
+      "step": 7100
+    },
+    {
+      "epoch": 10.36,
+      "learning_rate": 6.310534351145038e-05,
+      "loss": 2.038,
+      "step": 7200
+    },
+    {
+      "epoch": 10.5,
+      "learning_rate": 6.287633587786259e-05,
+      "loss": 2.0093,
+      "step": 7300
+    },
+    {
+      "epoch": 10.65,
+      "learning_rate": 6.26473282442748e-05,
+      "loss": 1.9839,
+      "step": 7400
+    },
+    {
+      "epoch": 10.79,
+      "learning_rate": 6.241832061068702e-05,
+      "loss": 1.9803,
+      "step": 7500
+    },
+    {
+      "epoch": 10.79,
+      "eval_cer": 0.2636592662572823,
+      "eval_loss": 0.7643126845359802,
+      "eval_runtime": 41.3574,
+      "eval_samples_per_second": 11.026,
+      "eval_steps_per_second": 1.378,
+      "eval_wer": 0.6562971911809121,
+      "step": 7500
+    },
+    {
+      "epoch": 10.93,
+      "learning_rate": 6.218931297709923e-05,
+      "loss": 1.9704,
+      "step": 7600
+    },
+    {
+      "epoch": 11.08,
+      "learning_rate": 6.196030534351144e-05,
+      "loss": 1.9923,
+      "step": 7700
+    },
+    {
+      "epoch": 11.22,
+      "learning_rate": 6.173129770992366e-05,
+      "loss": 1.9549,
+      "step": 7800
+    },
+    {
+      "epoch": 11.37,
+      "learning_rate": 6.150229007633587e-05,
+      "loss": 1.9339,
+      "step": 7900
+    },
+    {
+      "epoch": 11.51,
+      "learning_rate": 6.127328244274808e-05,
+      "loss": 1.9468,
+      "step": 8000
+    },
+    {
+      "epoch": 11.51,
+      "eval_cer": 0.25051173043615177,
+      "eval_loss": 0.7318933606147766,
+      "eval_runtime": 40.8808,
+      "eval_samples_per_second": 11.154,
+      "eval_steps_per_second": 1.394,
+      "eval_wer": 0.644065237088493,
+      "step": 8000
+    },
+    {
+      "epoch": 11.65,
+      "learning_rate": 6.10442748091603e-05,
+      "loss": 1.9691,
+      "step": 8100
+    },
+    {
+      "epoch": 11.8,
+      "learning_rate": 6.081526717557252e-05,
+      "loss": 1.9845,
+      "step": 8200
+    },
+    {
+      "epoch": 11.94,
+      "learning_rate": 6.0586259541984725e-05,
+      "loss": 1.9561,
+      "step": 8300
+    },
+    {
+      "epoch": 12.09,
+      "learning_rate": 6.035725190839694e-05,
+      "loss": 1.9486,
+      "step": 8400
+    },
+    {
+      "epoch": 12.23,
+      "learning_rate": 6.012824427480916e-05,
+      "loss": 1.9178,
+      "step": 8500
+    },
+    {
+      "epoch": 12.23,
+      "eval_cer": 0.24893717524799244,
+      "eval_loss": 0.6936821937561035,
+      "eval_runtime": 41.1829,
+      "eval_samples_per_second": 11.073,
+      "eval_steps_per_second": 1.384,
+      "eval_wer": 0.6319842947749924,
+      "step": 8500
+    },
+    {
+      "epoch": 12.37,
+      "learning_rate": 5.989923664122137e-05,
+      "loss": 1.9133,
+      "step": 8600
+    },
+    {
+      "epoch": 12.52,
+      "learning_rate": 5.9670229007633586e-05,
+      "loss": 1.9327,
+      "step": 8700
+    },
+    {
+      "epoch": 12.66,
+      "learning_rate": 5.944122137404579e-05,
+      "loss": 1.8749,
+      "step": 8800
+    },
+    {
+      "epoch": 12.8,
+      "learning_rate": 5.9212213740458006e-05,
+      "loss": 1.8775,
+      "step": 8900
+    },
+    {
+      "epoch": 12.95,
+      "learning_rate": 5.8983206106870226e-05,
+      "loss": 1.8515,
+      "step": 9000
+    },
+    {
+      "epoch": 12.95,
+      "eval_cer": 0.21961108486852465,
+      "eval_loss": 0.6443303823471069,
+      "eval_runtime": 40.5279,
+      "eval_samples_per_second": 11.252,
+      "eval_steps_per_second": 1.406,
+      "eval_wer": 0.6052552099063727,
+      "step": 9000
+    },
+    {
+      "epoch": 13.09,
+      "learning_rate": 5.875419847328244e-05,
+      "loss": 1.8554,
+      "step": 9100
+    },
+    {
+      "epoch": 13.24,
+      "learning_rate": 5.852519083969465e-05,
+      "loss": 1.8568,
+      "step": 9200
+    },
+    {
+      "epoch": 13.38,
+      "learning_rate": 5.829618320610686e-05,
+      "loss": 1.8477,
+      "step": 9300
+    },
+    {
+      "epoch": 13.52,
+      "learning_rate": 5.806717557251908e-05,
+      "loss": 1.8328,
+      "step": 9400
+    },
+    {
+      "epoch": 13.67,
+      "learning_rate": 5.783816793893129e-05,
+      "loss": 1.8083,
+      "step": 9500
+    },
+    {
+      "epoch": 13.67,
+      "eval_cer": 0.21484805542434263,
+      "eval_loss": 0.6285760402679443,
+      "eval_runtime": 41.6653,
+      "eval_samples_per_second": 10.944,
+      "eval_steps_per_second": 1.368,
+      "eval_wer": 0.6122017517366355,
+      "step": 9500
+    },
+    {
+      "epoch": 13.81,
+      "learning_rate": 5.760916030534351e-05,
+      "loss": 1.8236,
+      "step": 9600
+    },
+    {
+      "epoch": 13.96,
+      "learning_rate": 5.738015267175571e-05,
+      "loss": 1.8199,
+      "step": 9700
+    },
+    {
+      "epoch": 14.1,
+      "learning_rate": 5.7151145038167934e-05,
+      "loss": 1.8285,
+      "step": 9800
+    },
+    {
+      "epoch": 14.24,
+      "learning_rate": 5.692213740458015e-05,
+      "loss": 1.817,
+      "step": 9900
+    },
+    {
+      "epoch": 14.39,
+      "learning_rate": 5.669312977099236e-05,
+      "loss": 1.819,
+      "step": 10000
+    },
+    {
+      "epoch": 14.39,
+      "eval_cer": 0.2074476460399937,
+      "eval_loss": 0.6015097498893738,
+      "eval_runtime": 41.6458,
+      "eval_samples_per_second": 10.949,
+      "eval_steps_per_second": 1.369,
+      "eval_wer": 0.5986106916339474,
+      "step": 10000
+    },
+    {
+      "epoch": 14.53,
+      "learning_rate": 5.6466412213740455e-05,
+      "loss": 1.7952,
+      "step": 10100
+    },
+    {
+      "epoch": 14.68,
+      "learning_rate": 5.623740458015266e-05,
+      "loss": 1.7955,
+      "step": 10200
+    },
+    {
+      "epoch": 14.82,
+      "learning_rate": 5.600839694656488e-05,
+      "loss": 1.7878,
+      "step": 10300
+    },
+    {
+      "epoch": 14.96,
+      "learning_rate": 5.5779389312977095e-05,
+      "loss": 1.769,
+      "step": 10400
+    },
+    {
+      "epoch": 15.11,
+      "learning_rate": 5.555267175572519e-05,
+      "loss": 1.7684,
+      "step": 10500
+    },
+    {
+      "epoch": 15.11,
+      "eval_cer": 0.19815777042985355,
+      "eval_loss": 0.5682193636894226,
+      "eval_runtime": 41.2484,
+      "eval_samples_per_second": 11.055,
+      "eval_steps_per_second": 1.382,
+      "eval_wer": 0.5741467834491091,
+      "step": 10500
+    },
+    {
+      "epoch": 15.25,
+      "learning_rate": 5.53236641221374e-05,
+      "loss": 1.7626,
+      "step": 10600
+    },
+    {
+      "epoch": 15.4,
+      "learning_rate": 5.5094656488549616e-05,
+      "loss": 1.7582,
+      "step": 10700
+    },
+    {
+      "epoch": 15.54,
+      "learning_rate": 5.486564885496182e-05,
+      "loss": 1.75,
+      "step": 10800
+    },
+    {
+      "epoch": 15.68,
+      "learning_rate": 5.463664122137404e-05,
+      "loss": 1.735,
+      "step": 10900
+    },
+    {
+      "epoch": 15.83,
+      "learning_rate": 5.4407633587786256e-05,
+      "loss": 1.7195,
+      "step": 11000
+    },
+    {
+      "epoch": 15.83,
+      "eval_cer": 0.20067705873090852,
+      "eval_loss": 0.5385124683380127,
+      "eval_runtime": 41.6481,
+      "eval_samples_per_second": 10.949,
+      "eval_steps_per_second": 1.369,
+      "eval_wer": 0.5591966173361522,
+      "step": 11000
+    },
+    {
+      "epoch": 15.97,
+      "learning_rate": 5.417862595419847e-05,
+      "loss": 1.7274,
+      "step": 11100
+    },
+    {
+      "epoch": 16.11,
+      "learning_rate": 5.3949618320610677e-05,
+      "loss": 1.7183,
+      "step": 11200
+    },
+    {
+      "epoch": 16.26,
+      "learning_rate": 5.37206106870229e-05,
+      "loss": 1.7117,
+      "step": 11300
+    },
+    {
+      "epoch": 16.4,
+      "learning_rate": 5.349160305343511e-05,
+      "loss": 1.6918,
+      "step": 11400
+    },
+    {
+      "epoch": 16.55,
+      "learning_rate": 5.3262595419847324e-05,
+      "loss": 1.7044,
+      "step": 11500
+    },
+    {
+      "epoch": 16.55,
+      "eval_cer": 0.20965202330341678,
+      "eval_loss": 0.5361923575401306,
+      "eval_runtime": 41.5242,
+      "eval_samples_per_second": 10.982,
+      "eval_steps_per_second": 1.373,
+      "eval_wer": 0.5524010872848082,
+      "step": 11500
+    },
+    {
+      "epoch": 16.69,
+      "learning_rate": 5.303358778625954e-05,
+      "loss": 1.7134,
+      "step": 11600
+    },
+    {
+      "epoch": 16.83,
+      "learning_rate": 5.280458015267176e-05,
+      "loss": 1.7016,
+      "step": 11700
+    },
+    {
+      "epoch": 16.98,
+      "learning_rate": 5.2575572519083964e-05,
+      "loss": 1.7069,
+      "step": 11800
+    },
+    {
+      "epoch": 17.12,
+      "learning_rate": 5.234656488549618e-05,
+      "loss": 1.7046,
+      "step": 11900
+    },
+    {
+      "epoch": 17.27,
+      "learning_rate": 5.211755725190839e-05,
+      "loss": 1.6879,
+      "step": 12000
+    },
+    {
+      "epoch": 17.27,
+      "eval_cer": 0.20831365139348135,
+      "eval_loss": 0.5119141936302185,
+      "eval_runtime": 40.4618,
+      "eval_samples_per_second": 11.27,
+      "eval_steps_per_second": 1.409,
+      "eval_wer": 0.5489278163696768,
+      "step": 12000
+    },
+    {
+      "epoch": 17.41,
+      "learning_rate": 5.188854961832061e-05,
+      "loss": 1.681,
+      "step": 12100
+    },
+    {
+      "epoch": 17.55,
+      "learning_rate": 5.1659541984732825e-05,
+      "loss": 1.6683,
+      "step": 12200
+    },
+    {
+      "epoch": 17.7,
+      "learning_rate": 5.143053435114503e-05,
+      "loss": 1.655,
+      "step": 12300
+    },
+    {
+      "epoch": 17.84,
+      "learning_rate": 5.1201526717557245e-05,
+      "loss": 1.6604,
+      "step": 12400
+    },
+    {
+      "epoch": 17.98,
+      "learning_rate": 5.0972519083969465e-05,
+      "loss": 1.656,
+      "step": 12500
+    },
+    {
+      "epoch": 17.98,
+      "eval_cer": 0.19678003464021415,
+      "eval_loss": 0.4990406930446625,
+      "eval_runtime": 40.5826,
+      "eval_samples_per_second": 11.236,
+      "eval_steps_per_second": 1.405,
+      "eval_wer": 0.5362428269405014,
+      "step": 12500
+    },
+    {
+      "epoch": 18.13,
+      "learning_rate": 5.074351145038168e-05,
+      "loss": 1.6645,
+      "step": 12600
+    },
+    {
+      "epoch": 18.27,
+      "learning_rate": 5.051450381679389e-05,
+      "loss": 1.6269,
+      "step": 12700
+    },
+    {
+      "epoch": 18.42,
+      "learning_rate": 5.02854961832061e-05,
+      "loss": 1.6306,
+      "step": 12800
+    },
+    {
+      "epoch": 18.56,
+      "learning_rate": 5.005877862595419e-05,
+      "loss": 1.6191,
+      "step": 12900
+    },
+    {
+      "epoch": 18.7,
+      "learning_rate": 4.9829770992366406e-05,
+      "loss": 1.6122,
+      "step": 13000
+    },
+    {
+      "epoch": 18.7,
+      "eval_cer": 0.18997008345142496,
+      "eval_loss": 0.45614466071128845,
+      "eval_runtime": 41.2927,
+      "eval_samples_per_second": 11.043,
+      "eval_steps_per_second": 1.38,
+      "eval_wer": 0.5092117185140441,
+      "step": 13000
+    },
+    {
+      "epoch": 18.85,
+      "learning_rate": 4.9600763358778626e-05,
+      "loss": 1.622,
+      "step": 13100
+    },
+    {
+      "epoch": 18.99,
+      "learning_rate": 4.937175572519084e-05,
+      "loss": 1.6305,
+      "step": 13200
+    },
+    {
+      "epoch": 19.14,
+      "learning_rate": 4.9142748091603046e-05,
+      "loss": 1.6134,
+      "step": 13300
+    },
+    {
+      "epoch": 19.28,
+      "learning_rate": 4.891374045801526e-05,
+      "loss": 1.6044,
+      "step": 13400
+    },
+    {
+      "epoch": 19.42,
+      "learning_rate": 4.868473282442748e-05,
+      "loss": 1.5919,
+      "step": 13500
+    },
+    {
+      "epoch": 19.42,
+      "eval_cer": 0.19752794835458984,
+      "eval_loss": 0.47778981924057007,
+      "eval_runtime": 41.5758,
+      "eval_samples_per_second": 10.968,
+      "eval_steps_per_second": 1.371,
+      "eval_wer": 0.5225007550588946,
+      "step": 13500
+    },
+    {
+      "epoch": 19.57,
+      "learning_rate": 4.8455725190839694e-05,
+      "loss": 1.595,
+      "step": 13600
+    },
+    {
+      "epoch": 19.71,
+      "learning_rate": 4.822671755725191e-05,
+      "loss": 1.5959,
+      "step": 13700
+    },
+    {
+      "epoch": 19.86,
+      "learning_rate": 4.7997709923664114e-05,
+      "loss": 1.6006,
+      "step": 13800
+    },
+    {
+      "epoch": 20.0,
+      "learning_rate": 4.7768702290076334e-05,
+      "loss": 1.5913,
+      "step": 13900
+    },
+    {
+      "epoch": 20.14,
+      "learning_rate": 4.753969465648855e-05,
+      "loss": 1.5896,
+      "step": 14000
+    },
+    {
+      "epoch": 20.14,
+      "eval_cer": 0.18591560384191466,
+      "eval_loss": 0.4563109278678894,
+      "eval_runtime": 40.8794,
+      "eval_samples_per_second": 11.155,
+      "eval_steps_per_second": 1.394,
+      "eval_wer": 0.5098157656297191,
+      "step": 14000
+    },
+    {
+      "epoch": 20.29,
+      "learning_rate": 4.731068702290076e-05,
+      "loss": 1.5823,
+      "step": 14100
+    },
+    {
+      "epoch": 20.43,
+      "learning_rate": 4.708167938931297e-05,
+      "loss": 1.5634,
+      "step": 14200
+    },
+    {
+      "epoch": 20.57,
+      "learning_rate": 4.685267175572519e-05,
+      "loss": 1.5573,
+      "step": 14300
+    },
+    {
+      "epoch": 20.72,
+      "learning_rate": 4.66236641221374e-05,
+      "loss": 1.5689,
+      "step": 14400
+    },
+    {
+      "epoch": 20.86,
+      "learning_rate": 4.6394656488549615e-05,
+      "loss": 1.5589,
+      "step": 14500
+    },
+    {
+      "epoch": 20.86,
+      "eval_cer": 0.17249252086285624,
+      "eval_loss": 0.43622052669525146,
+      "eval_runtime": 41.7277,
+      "eval_samples_per_second": 10.928,
+      "eval_steps_per_second": 1.366,
+      "eval_wer": 0.4939595288432498,
+      "step": 14500
+    },
+    {
+      "epoch": 21.01,
+      "learning_rate": 4.616564885496183e-05,
+      "loss": 1.5697,
+      "step": 14600
+    },
+    {
+      "epoch": 21.15,
+      "learning_rate": 4.593664122137405e-05,
+      "loss": 1.5336,
+      "step": 14700
+    },
+    {
+      "epoch": 21.29,
+      "learning_rate": 4.5707633587786255e-05,
+      "loss": 1.5425,
+      "step": 14800
+    },
+    {
+      "epoch": 21.44,
+      "learning_rate": 4.547862595419847e-05,
+      "loss": 1.5461,
+      "step": 14900
+    },
+    {
+      "epoch": 21.58,
+      "learning_rate": 4.524961832061068e-05,
+      "loss": 1.5353,
+      "step": 15000
+    },
+    {
+      "epoch": 21.58,
+      "eval_cer": 0.15804597701149425,
+      "eval_loss": 0.41395294666290283,
+      "eval_runtime": 40.5257,
+      "eval_samples_per_second": 11.252,
+      "eval_steps_per_second": 1.407,
+      "eval_wer": 0.4826336454243431,
+      "step": 15000
+    },
+    {
+      "epoch": 21.73,
+      "learning_rate": 4.5020610687022895e-05,
+      "loss": 1.5348,
+      "step": 15100
+    },
+    {
+      "epoch": 21.87,
+      "learning_rate": 4.4791603053435116e-05,
+      "loss": 1.5279,
+      "step": 15200
+    },
+    {
+      "epoch": 22.01,
+      "learning_rate": 4.456259541984732e-05,
+      "loss": 1.5492,
+      "step": 15300
+    },
+    {
+      "epoch": 22.16,
+      "learning_rate": 4.4333587786259536e-05,
+      "loss": 1.5291,
+      "step": 15400
+    },
+    {
+      "epoch": 22.3,
+      "learning_rate": 4.410458015267175e-05,
+      "loss": 1.5441,
+      "step": 15500
+    },
+    {
+      "epoch": 22.3,
+      "eval_cer": 0.15501495827428752,
+      "eval_loss": 0.40313535928726196,
+      "eval_runtime": 41.0848,
+      "eval_samples_per_second": 11.099,
+      "eval_steps_per_second": 1.387,
+      "eval_wer": 0.47417698580489276,
+      "step": 15500
+    },
+    {
+      "epoch": 22.45,
+      "learning_rate": 4.387557251908397e-05,
+      "loss": 1.518,
+      "step": 15600
+    },
+    {
+      "epoch": 22.59,
+      "learning_rate": 4.364656488549618e-05,
+      "loss": 1.5081,
+      "step": 15700
+    },
+    {
+      "epoch": 22.73,
+      "learning_rate": 4.341755725190839e-05,
+      "loss": 1.4959,
+      "step": 15800
+    },
+    {
+      "epoch": 22.88,
+      "learning_rate": 4.31885496183206e-05,
+      "loss": 1.5097,
+      "step": 15900
+    },
+    {
+      "epoch": 23.02,
+      "learning_rate": 4.295954198473282e-05,
+      "loss": 1.5116,
+      "step": 16000
+    },
+    {
+      "epoch": 23.02,
+      "eval_cer": 0.15450322783813572,
+      "eval_loss": 0.39162585139274597,
+      "eval_runtime": 40.7373,
+      "eval_samples_per_second": 11.194,
+      "eval_steps_per_second": 1.399,
+      "eval_wer": 0.4747810329205678,
+      "step": 16000
+    },
+    {
+      "epoch": 23.17,
+      "learning_rate": 4.273053435114504e-05,
+      "loss": 1.4951,
+      "step": 16100
+    },
+    {
+      "epoch": 23.31,
+      "learning_rate": 4.250152671755724e-05,
+      "loss": 1.4974,
+      "step": 16200
+    },
+    {
+      "epoch": 23.45,
+      "learning_rate": 4.227480916030534e-05,
+      "loss": 1.5045,
+      "step": 16300
+    },
+    {
+      "epoch": 23.6,
+      "learning_rate": 4.204580152671755e-05,
+      "loss": 1.4944,
+      "step": 16400
+    },
+    {
+      "epoch": 23.74,
+      "learning_rate": 4.181679389312977e-05,
+      "loss": 1.4731,
+      "step": 16500
+    },
+    {
+      "epoch": 23.74,
+      "eval_cer": 0.15422768068020784,
+      "eval_loss": 0.3840835392475128,
+      "eval_runtime": 40.8763,
+      "eval_samples_per_second": 11.156,
+      "eval_steps_per_second": 1.394,
+      "eval_wer": 0.4809725158562368,
+      "step": 16500
+    },
+    {
+      "epoch": 23.88,
+      "learning_rate": 4.1587786259541985e-05,
+      "loss": 1.472,
+      "step": 16600
+    },
+    {
+      "epoch": 24.03,
+      "learning_rate": 4.13587786259542e-05,
+      "loss": 1.4847,
+      "step": 16700
+    },
+    {
+      "epoch": 24.17,
+      "learning_rate": 4.1129770992366405e-05,
+      "loss": 1.4603,
+      "step": 16800
+    },
+    {
+      "epoch": 24.32,
+      "learning_rate": 4.090076335877862e-05,
+      "loss": 1.4563,
+      "step": 16900
+    },
+    {
+      "epoch": 24.46,
+      "learning_rate": 4.067175572519084e-05,
+      "loss": 1.4647,
+      "step": 17000
+    },
+    {
+      "epoch": 24.46,
+      "eval_cer": 0.14753582113053063,
+      "eval_loss": 0.37518319487571716,
+      "eval_runtime": 41.0205,
+      "eval_samples_per_second": 11.116,
+      "eval_steps_per_second": 1.39,
+      "eval_wer": 0.452431289640592,
+      "step": 17000
+    },
+    {
+      "epoch": 24.6,
+      "learning_rate": 4.044274809160305e-05,
+      "loss": 1.4585,
+      "step": 17100
+    },
+    {
+      "epoch": 24.75,
+      "learning_rate": 4.021374045801526e-05,
+      "loss": 1.4692,
+      "step": 17200
+    },
+    {
+      "epoch": 24.89,
+      "learning_rate": 3.998473282442747e-05,
+      "loss": 1.444,
+      "step": 17300
+    },
+    {
+      "epoch": 25.04,
+      "learning_rate": 3.975572519083969e-05,
+      "loss": 1.4717,
+      "step": 17400
+    },
+    {
+      "epoch": 25.18,
+      "learning_rate": 3.9526717557251906e-05,
+      "loss": 1.4328,
+      "step": 17500
+    },
+    {
+      "epoch": 25.18,
+      "eval_cer": 0.1461187214611872,
+      "eval_loss": 0.35870596766471863,
+      "eval_runtime": 40.6723,
+      "eval_samples_per_second": 11.212,
+      "eval_steps_per_second": 1.401,
+      "eval_wer": 0.4475989127151918,
+      "step": 17500
+    },
+    {
+      "epoch": 25.32,
+      "learning_rate": 3.929770992366412e-05,
+      "loss": 1.4329,
+      "step": 17600
+    },
+    {
+      "epoch": 25.47,
+      "learning_rate": 3.9068702290076326e-05,
+      "loss": 1.4209,
+      "step": 17700
+    },
+    {
+      "epoch": 25.61,
+      "learning_rate": 3.884198473282442e-05,
+      "loss": 1.4188,
+      "step": 17800
+    },
+    {
+      "epoch": 25.75,
+      "learning_rate": 3.861297709923664e-05,
+      "loss": 1.4301,
+      "step": 17900
+    },
+    {
+      "epoch": 25.9,
+      "learning_rate": 3.8383969465648854e-05,
+      "loss": 1.4129,
+      "step": 18000
+    },
+    {
+      "epoch": 25.9,
+      "eval_cer": 0.13663202645252717,
+      "eval_loss": 0.3428773581981659,
+      "eval_runtime": 42.0192,
+      "eval_samples_per_second": 10.852,
+      "eval_steps_per_second": 1.357,
+      "eval_wer": 0.42419208698278466,
+      "step": 18000
+    },
+    {
+      "epoch": 26.04,
+      "learning_rate": 3.815496183206107e-05,
+      "loss": 1.4266,
+      "step": 18100
+    },
+    {
+      "epoch": 26.19,
+      "learning_rate": 3.7925954198473274e-05,
+      "loss": 1.4166,
+      "step": 18200
+    },
+    {
+      "epoch": 26.33,
+      "learning_rate": 3.7696946564885494e-05,
+      "loss": 1.4157,
+      "step": 18300
+    },
+    {
+      "epoch": 26.47,
+      "learning_rate": 3.746793893129771e-05,
+      "loss": 1.4285,
+      "step": 18400
+    },
+    {
+      "epoch": 26.62,
+      "learning_rate": 3.723893129770992e-05,
+      "loss": 1.4062,
+      "step": 18500
+    },
+    {
+      "epoch": 26.62,
+      "eval_cer": 0.13549047394111163,
+      "eval_loss": 0.34499478340148926,
+      "eval_runtime": 41.0336,
+      "eval_samples_per_second": 11.113,
+      "eval_steps_per_second": 1.389,
+      "eval_wer": 0.4250981576562972,
+      "step": 18500
+    },
+    {
+      "epoch": 26.76,
+      "learning_rate": 3.7009923664122134e-05,
+      "loss": 1.4163,
+      "step": 18600
+    },
+    {
+      "epoch": 26.91,
+      "learning_rate": 3.678091603053435e-05,
+      "loss": 1.404,
+      "step": 18700
+    },
+    {
+      "epoch": 27.05,
+      "learning_rate": 3.655190839694656e-05,
+      "loss": 1.4134,
+      "step": 18800
+    },
+    {
+      "epoch": 27.19,
+      "learning_rate": 3.6322900763358775e-05,
+      "loss": 1.4001,
+      "step": 18900
+    },
+    {
+      "epoch": 27.34,
+      "learning_rate": 3.609389312977099e-05,
+      "loss": 1.3928,
+      "step": 19000
+    },
+    {
+      "epoch": 27.34,
+      "eval_cer": 0.13218390804597702,
+      "eval_loss": 0.32969579100608826,
+      "eval_runtime": 41.0801,
+      "eval_samples_per_second": 11.1,
+      "eval_steps_per_second": 1.388,
+      "eval_wer": 0.4145273331319843,
+      "step": 19000
+    },
+    {
+      "epoch": 27.48,
+      "learning_rate": 3.58648854961832e-05,
+      "loss": 1.3979,
+      "step": 19100
+    },
+    {
+      "epoch": 27.63,
+      "learning_rate": 3.5635877862595415e-05,
+      "loss": 1.3971,
+      "step": 19200
+    },
+    {
+      "epoch": 27.77,
+      "learning_rate": 3.540687022900763e-05,
+      "loss": 1.3934,
+      "step": 19300
+    },
+    {
+      "epoch": 27.91,
+      "learning_rate": 3.517786259541984e-05,
+      "loss": 1.3866,
+      "step": 19400
+    },
+    {
+      "epoch": 28.06,
+      "learning_rate": 3.4948854961832055e-05,
+      "loss": 1.3906,
+      "step": 19500
+    },
+    {
+      "epoch": 28.06,
+      "eval_cer": 0.1336403715950244,
+      "eval_loss": 0.32101842761039734,
+      "eval_runtime": 41.0367,
+      "eval_samples_per_second": 11.112,
+      "eval_steps_per_second": 1.389,
+      "eval_wer": 0.4184536393838719,
+      "step": 19500
+    },
+    {
+      "epoch": 28.2,
+      "learning_rate": 3.471984732824427e-05,
+      "loss": 1.3689,
+      "step": 19600
+    },
+    {
+      "epoch": 28.34,
+      "learning_rate": 3.449083969465649e-05,
+      "loss": 1.3715,
+      "step": 19700
+    },
+    {
+      "epoch": 28.49,
+      "learning_rate": 3.4261832061068696e-05,
+      "loss": 1.3527,
+      "step": 19800
+    },
+    {
+      "epoch": 28.63,
+      "learning_rate": 3.4032824427480916e-05,
+      "loss": 1.3532,
+      "step": 19900
+    },
+    {
+      "epoch": 28.78,
+      "learning_rate": 3.380381679389312e-05,
+      "loss": 1.358,
+      "step": 20000
+    },
+    {
+      "epoch": 28.78,
+      "eval_cer": 0.12753897024090693,
+      "eval_loss": 0.31306591629981995,
+      "eval_runtime": 41.2359,
+      "eval_samples_per_second": 11.058,
+      "eval_steps_per_second": 1.382,
+      "eval_wer": 0.39700996677740863,
+      "step": 20000
+    },
+    {
+      "epoch": 28.92,
+      "learning_rate": 3.357480916030534e-05,
+      "loss": 1.3582,
+      "step": 20100
+    },
+    {
+      "epoch": 29.06,
+      "learning_rate": 3.334580152671755e-05,
+      "loss": 1.3587,
+      "step": 20200
+    },
+    {
+      "epoch": 29.21,
+      "learning_rate": 3.311679389312977e-05,
+      "loss": 1.3392,
+      "step": 20300
+    },
+    {
+      "epoch": 29.35,
+      "learning_rate": 3.288778625954198e-05,
+      "loss": 1.3486,
+      "step": 20400
+    },
+    {
+      "epoch": 29.5,
+      "learning_rate": 3.26587786259542e-05,
+      "loss": 1.3445,
+      "step": 20500
+    },
+    {
+      "epoch": 29.5,
+      "eval_cer": 0.12761769800031492,
+      "eval_loss": 0.3069218099117279,
+      "eval_runtime": 41.0687,
+      "eval_samples_per_second": 11.103,
+      "eval_steps_per_second": 1.388,
+      "eval_wer": 0.3920265780730897,
+      "step": 20500
+    },
+    {
+      "epoch": 29.64,
+      "learning_rate": 3.242977099236641e-05,
+      "loss": 1.3354,
+      "step": 20600
+    },
+    {
+      "epoch": 29.78,
+      "learning_rate": 3.2200763358778624e-05,
+      "loss": 1.3334,
+      "step": 20700
+    },
+    {
+      "epoch": 29.93,
+      "learning_rate": 3.197175572519084e-05,
+      "loss": 1.3305,
+      "step": 20800
+    },
+    {
+      "epoch": 30.07,
+      "learning_rate": 3.174274809160305e-05,
+      "loss": 1.3354,
+      "step": 20900
+    },
+    {
+      "epoch": 30.22,
+      "learning_rate": 3.1513740458015264e-05,
+      "loss": 1.3159,
+      "step": 21000
+    },
+    {
+      "epoch": 30.22,
+      "eval_cer": 0.1254920484962998,
+      "eval_loss": 0.30346596240997314,
+      "eval_runtime": 41.0784,
+      "eval_samples_per_second": 11.101,
+      "eval_steps_per_second": 1.388,
+      "eval_wer": 0.3961038961038961,
+      "step": 21000
+    },
+    {
+      "epoch": 30.36,
+      "learning_rate": 3.128473282442748e-05,
+      "loss": 1.3376,
+      "step": 21100
+    },
+    {
+      "epoch": 30.5,
+      "learning_rate": 3.105572519083969e-05,
+      "loss": 1.324,
+      "step": 21200
+    },
+    {
+      "epoch": 30.65,
+      "learning_rate": 3.0826717557251904e-05,
+      "loss": 1.3091,
+      "step": 21300
+    },
+    {
+      "epoch": 30.79,
+      "learning_rate": 3.059770992366412e-05,
+      "loss": 1.3213,
+      "step": 21400
+    },
+    {
+      "epoch": 30.93,
+      "learning_rate": 3.0368702290076335e-05,
+      "loss": 1.3044,
+      "step": 21500
+    },
+    {
+      "epoch": 30.93,
+      "eval_cer": 0.12423240434577232,
+      "eval_loss": 0.29519879817962646,
+      "eval_runtime": 41.1753,
+      "eval_samples_per_second": 11.075,
+      "eval_steps_per_second": 1.384,
+      "eval_wer": 0.3853820598006645,
+      "step": 21500
+    },
+    {
+      "epoch": 31.08,
+      "learning_rate": 3.0139694656488545e-05,
+      "loss": 1.3033,
+      "step": 21600
+    },
+    {
+      "epoch": 31.22,
+      "learning_rate": 2.991068702290076e-05,
+      "loss": 1.2995,
+      "step": 21700
+    },
+    {
+      "epoch": 31.37,
+      "learning_rate": 2.9681679389312975e-05,
+      "loss": 1.3101,
+      "step": 21800
+    },
+    {
+      "epoch": 31.51,
+      "learning_rate": 2.945267175572519e-05,
+      "loss": 1.304,
+      "step": 21900
+    },
+    {
+      "epoch": 31.65,
+      "learning_rate": 2.9223664122137402e-05,
+      "loss": 1.3034,
+      "step": 22000
+    },
+    {
+      "epoch": 31.65,
+      "eval_cer": 0.12273657691702095,
+      "eval_loss": 0.29660850763320923,
+      "eval_runtime": 41.8912,
+      "eval_samples_per_second": 10.885,
+      "eval_steps_per_second": 1.361,
+      "eval_wer": 0.37722742373905166,
+      "step": 22000
+    },
+    {
+      "epoch": 31.8,
+      "learning_rate": 2.8994656488549615e-05,
+      "loss": 1.2912,
+      "step": 22100
+    },
+    {
+      "epoch": 31.94,
+      "learning_rate": 2.876564885496183e-05,
+      "loss": 1.299,
+      "step": 22200
+    },
+    {
+      "epoch": 32.09,
+      "learning_rate": 2.8536641221374046e-05,
+      "loss": 1.3042,
+      "step": 22300
+    },
+    {
+      "epoch": 32.23,
+      "learning_rate": 2.8307633587786256e-05,
+      "loss": 1.294,
+      "step": 22400
+    },
+    {
+      "epoch": 32.37,
+      "learning_rate": 2.8078625954198472e-05,
+      "loss": 1.2963,
+      "step": 22500
+    },
+    {
+      "epoch": 32.37,
+      "eval_cer": 0.12080774681152574,
+      "eval_loss": 0.2843906879425049,
+      "eval_runtime": 41.7644,
+      "eval_samples_per_second": 10.918,
+      "eval_steps_per_second": 1.365,
+      "eval_wer": 0.3705829054666264,
+      "step": 22500
+    },
+    {
+      "epoch": 32.52,
+      "learning_rate": 2.7849618320610682e-05,
+      "loss": 1.2769,
+      "step": 22600
+    },
+    {
+      "epoch": 32.66,
+      "learning_rate": 2.76206106870229e-05,
+      "loss": 1.2812,
+      "step": 22700
+    },
+    {
+      "epoch": 32.8,
+      "learning_rate": 2.7391603053435113e-05,
+      "loss": 1.2827,
+      "step": 22800
+    },
+    {
+      "epoch": 32.95,
+      "learning_rate": 2.7162595419847326e-05,
+      "loss": 1.2747,
+      "step": 22900
+    },
+    {
+      "epoch": 33.09,
+      "learning_rate": 2.6935877862595417e-05,
+      "loss": 1.2765,
+      "step": 23000
+    },
+    {
+      "epoch": 33.09,
+      "eval_cer": 0.11726499763816722,
+      "eval_loss": 0.28407707810401917,
+      "eval_runtime": 40.9894,
+      "eval_samples_per_second": 11.125,
+      "eval_steps_per_second": 1.391,
+      "eval_wer": 0.35668982180610087,
+      "step": 23000
+    },
+    {
+      "epoch": 33.24,
+      "learning_rate": 2.670687022900763e-05,
+      "loss": 1.2785,
+      "step": 23100
+    },
+    {
+      "epoch": 33.38,
+      "learning_rate": 2.6477862595419844e-05,
+      "loss": 1.2644,
+      "step": 23200
+    },
+    {
+      "epoch": 33.52,
+      "learning_rate": 2.624885496183206e-05,
+      "loss": 1.2724,
+      "step": 23300
+    },
+    {
+      "epoch": 33.67,
+      "learning_rate": 2.601984732824427e-05,
+      "loss": 1.2551,
+      "step": 23400
+    },
+    {
+      "epoch": 33.81,
+      "learning_rate": 2.5790839694656488e-05,
+      "loss": 1.2438,
+      "step": 23500
+    },
+    {
+      "epoch": 33.81,
+      "eval_cer": 0.11372224846480869,
+      "eval_loss": 0.2734295129776001,
+      "eval_runtime": 41.7199,
+      "eval_samples_per_second": 10.93,
+      "eval_steps_per_second": 1.366,
+      "eval_wer": 0.35517970401691334,
+      "step": 23500
+    },
+    {
+      "epoch": 33.96,
+      "learning_rate": 2.5561832061068698e-05,
+      "loss": 1.2491,
+      "step": 23600
+    },
+    {
+      "epoch": 34.1,
+      "learning_rate": 2.5332824427480915e-05,
+      "loss": 1.252,
+      "step": 23700
+    },
+    {
+      "epoch": 34.24,
+      "learning_rate": 2.5103816793893128e-05,
+      "loss": 1.2467,
+      "step": 23800
+    },
+    {
+      "epoch": 34.39,
+      "learning_rate": 2.487480916030534e-05,
+      "loss": 1.2406,
+      "step": 23900
+    },
+    {
+      "epoch": 34.53,
+      "learning_rate": 2.4645801526717555e-05,
+      "loss": 1.2487,
+      "step": 24000
+    },
+    {
+      "epoch": 34.53,
+      "eval_cer": 0.11179341835931349,
+      "eval_loss": 0.2702818512916565,
+      "eval_runtime": 41.8515,
+      "eval_samples_per_second": 10.896,
+      "eval_steps_per_second": 1.362,
+      "eval_wer": 0.3501963153125944,
+      "step": 24000
+    },
+    {
+      "epoch": 34.68,
+      "learning_rate": 2.441679389312977e-05,
+      "loss": 1.2504,
+      "step": 24100
+    },
+    {
+      "epoch": 34.82,
+      "learning_rate": 2.4187786259541982e-05,
+      "loss": 1.2341,
+      "step": 24200
+    },
+    {
+      "epoch": 34.96,
+      "learning_rate": 2.39587786259542e-05,
+      "loss": 1.2477,
+      "step": 24300
+    },
+    {
+      "epoch": 35.11,
+      "learning_rate": 2.372977099236641e-05,
+      "loss": 1.2427,
+      "step": 24400
+    },
+    {
+      "epoch": 35.25,
+      "learning_rate": 2.3500763358778626e-05,
+      "loss": 1.2249,
+      "step": 24500
+    },
+    {
+      "epoch": 35.25,
+      "eval_cer": 0.11423397890096047,
+      "eval_loss": 0.2650163471698761,
+      "eval_runtime": 41.2103,
+      "eval_samples_per_second": 11.065,
+      "eval_steps_per_second": 1.383,
+      "eval_wer": 0.3483841739655693,
+      "step": 24500
+    },
+    {
+      "epoch": 35.4,
+      "learning_rate": 2.3271755725190836e-05,
+      "loss": 1.2265,
+      "step": 24600
+    },
+    {
+      "epoch": 35.54,
+      "learning_rate": 2.3042748091603052e-05,
+      "loss": 1.2276,
+      "step": 24700
+    },
+    {
+      "epoch": 35.68,
+      "learning_rate": 2.2816030534351143e-05,
+      "loss": 1.2332,
+      "step": 24800
+    },
+    {
+      "epoch": 35.83,
+      "learning_rate": 2.2587022900763357e-05,
+      "loss": 1.2249,
+      "step": 24900
+    },
+    {
+      "epoch": 35.97,
+      "learning_rate": 2.235801526717557e-05,
+      "loss": 1.2229,
+      "step": 25000
+    },
+    {
+      "epoch": 35.97,
+      "eval_cer": 0.10970713273500236,
+      "eval_loss": 0.25843024253845215,
+      "eval_runtime": 42.815,
+      "eval_samples_per_second": 10.65,
+      "eval_steps_per_second": 1.331,
+      "eval_wer": 0.3373603141045001,
+      "step": 25000
+    },
+    {
+      "epoch": 36.11,
+      "learning_rate": 2.2129007633587784e-05,
+      "loss": 1.2412,
+      "step": 25100
+    },
+    {
+      "epoch": 36.26,
+      "learning_rate": 2.1899999999999997e-05,
+      "loss": 1.212,
+      "step": 25200
+    },
+    {
+      "epoch": 36.4,
+      "learning_rate": 2.1670992366412214e-05,
+      "loss": 1.2151,
+      "step": 25300
+    },
+    {
+      "epoch": 36.55,
+      "learning_rate": 2.1441984732824424e-05,
+      "loss": 1.2303,
+      "step": 25400
+    },
+    {
+      "epoch": 36.69,
+      "learning_rate": 2.121297709923664e-05,
+      "loss": 1.2374,
+      "step": 25500
+    },
+    {
+      "epoch": 36.69,
+      "eval_cer": 0.10951031333648244,
+      "eval_loss": 0.2568279504776001,
+      "eval_runtime": 41.6839,
+      "eval_samples_per_second": 10.939,
+      "eval_steps_per_second": 1.367,
+      "eval_wer": 0.33373603141045,
+      "step": 25500
+    },
+    {
+      "epoch": 36.83,
+      "learning_rate": 2.098396946564885e-05,
+      "loss": 1.2152,
+      "step": 25600
+    },
+    {
+      "epoch": 36.98,
+      "learning_rate": 2.0754961832061068e-05,
+      "loss": 1.2089,
+      "step": 25700
+    },
+    {
+      "epoch": 37.12,
+      "learning_rate": 2.052595419847328e-05,
+      "loss": 1.2201,
+      "step": 25800
+    },
+    {
+      "epoch": 37.27,
+      "learning_rate": 2.0296946564885495e-05,
+      "loss": 1.2006,
+      "step": 25900
+    },
+    {
+      "epoch": 37.41,
+      "learning_rate": 2.0067938931297708e-05,
+      "loss": 1.2153,
+      "step": 26000
+    },
+    {
+      "epoch": 37.41,
+      "eval_cer": 0.10710911667453944,
+      "eval_loss": 0.24941784143447876,
+      "eval_runtime": 41.3494,
+      "eval_samples_per_second": 11.028,
+      "eval_steps_per_second": 1.378,
+      "eval_wer": 0.33267894895801875,
+      "step": 26000
+    },
+    {
+      "epoch": 37.55,
+      "learning_rate": 1.983893129770992e-05,
+      "loss": 1.2071,
+      "step": 26100
+    },
+    {
+      "epoch": 37.7,
+      "learning_rate": 1.9609923664122135e-05,
+      "loss": 1.2042,
+      "step": 26200
+    },
+    {
+      "epoch": 37.84,
+      "learning_rate": 1.9380916030534352e-05,
+      "loss": 1.2037,
+      "step": 26300
+    },
+    {
+      "epoch": 37.98,
+      "learning_rate": 1.9151908396946562e-05,
+      "loss": 1.1962,
+      "step": 26400
+    },
+    {
+      "epoch": 38.13,
+      "learning_rate": 1.892290076335878e-05,
+      "loss": 1.1925,
+      "step": 26500
+    },
+    {
+      "epoch": 38.13,
+      "eval_cer": 0.1076995748700992,
+      "eval_loss": 0.2518324553966522,
+      "eval_runtime": 40.748,
+      "eval_samples_per_second": 11.191,
+      "eval_steps_per_second": 1.399,
+      "eval_wer": 0.33660525520990636,
+      "step": 26500
+    },
+    {
+      "epoch": 38.27,
+      "learning_rate": 1.869389312977099e-05,
+      "loss": 1.1969,
+      "step": 26600
+    },
+    {
+      "epoch": 38.42,
+      "learning_rate": 1.8464885496183202e-05,
+      "loss": 1.1947,
+      "step": 26700
+    },
+    {
+      "epoch": 38.56,
+      "learning_rate": 1.823587786259542e-05,
+      "loss": 1.2005,
+      "step": 26800
+    },
+    {
+      "epoch": 38.7,
+      "learning_rate": 1.8006870229007632e-05,
+      "loss": 1.1961,
+      "step": 26900
+    },
+    {
+      "epoch": 38.85,
+      "learning_rate": 1.7777862595419846e-05,
+      "loss": 1.1908,
+      "step": 27000
+    },
+    {
+      "epoch": 38.85,
+      "eval_cer": 0.10565265312549205,
+      "eval_loss": 0.24367305636405945,
+      "eval_runtime": 41.2308,
+      "eval_samples_per_second": 11.06,
+      "eval_steps_per_second": 1.382,
+      "eval_wer": 0.3272425249169435,
+      "step": 27000
+    },
+    {
+      "epoch": 38.99,
+      "learning_rate": 1.754885496183206e-05,
+      "loss": 1.1762,
+      "step": 27100
+    },
+    {
+      "epoch": 39.14,
+      "learning_rate": 1.7319847328244273e-05,
+      "loss": 1.2018,
+      "step": 27200
+    },
+    {
+      "epoch": 39.28,
+      "learning_rate": 1.7090839694656486e-05,
+      "loss": 1.1822,
+      "step": 27300
+    },
+    {
+      "epoch": 39.42,
+      "learning_rate": 1.68618320610687e-05,
+      "loss": 1.1745,
+      "step": 27400
+    },
+    {
+      "epoch": 39.57,
+      "learning_rate": 1.6632824427480913e-05,
+      "loss": 1.1858,
+      "step": 27500
+    },
+    {
+      "epoch": 39.57,
+      "eval_cer": 0.10443237285466855,
+      "eval_loss": 0.23960824310779572,
+      "eval_runtime": 42.517,
+      "eval_samples_per_second": 10.725,
+      "eval_steps_per_second": 1.341,
+      "eval_wer": 0.32648746602234974,
+      "step": 27500
+    },
+    {
+      "epoch": 39.71,
+      "learning_rate": 1.6403816793893127e-05,
+      "loss": 1.1866,
+      "step": 27600
+    },
+    {
+      "epoch": 39.86,
+      "learning_rate": 1.617480916030534e-05,
+      "loss": 1.1878,
+      "step": 27700
+    },
+    {
+      "epoch": 40.0,
+      "learning_rate": 1.5945801526717557e-05,
+      "loss": 1.1817,
+      "step": 27800
+    },
+    {
+      "epoch": 40.14,
+      "learning_rate": 1.571679389312977e-05,
+      "loss": 1.1851,
+      "step": 27900
+    },
+    {
+      "epoch": 40.29,
+      "learning_rate": 1.5487786259541984e-05,
+      "loss": 1.1808,
+      "step": 28000
+    },
+    {
+      "epoch": 40.29,
+      "eval_cer": 0.10277908990710125,
+      "eval_loss": 0.2373325228691101,
+      "eval_runtime": 41.3513,
+      "eval_samples_per_second": 11.027,
+      "eval_steps_per_second": 1.378,
+      "eval_wer": 0.31561461794019935,
+      "step": 28000
+    },
+    {
+      "epoch": 40.43,
+      "learning_rate": 1.5258778625954197e-05,
+      "loss": 1.1558,
+      "step": 28100
+    },
+    {
+      "epoch": 40.57,
+      "learning_rate": 1.502977099236641e-05,
+      "loss": 1.1804,
+      "step": 28200
+    },
+    {
+      "epoch": 40.72,
+      "learning_rate": 1.4800763358778624e-05,
+      "loss": 1.1736,
+      "step": 28300
+    },
+    {
+      "epoch": 40.86,
+      "learning_rate": 1.4571755725190838e-05,
+      "loss": 1.1782,
+      "step": 28400
+    },
+    {
+      "epoch": 41.01,
+      "learning_rate": 1.4342748091603053e-05,
+      "loss": 1.1842,
+      "step": 28500
+    },
+    {
+      "epoch": 41.01,
+      "eval_cer": 0.10258227050858132,
+      "eval_loss": 0.23562349379062653,
+      "eval_runtime": 40.492,
+      "eval_samples_per_second": 11.261,
+      "eval_steps_per_second": 1.408,
+      "eval_wer": 0.31516158260344307,
+      "step": 28500
+    },
+    {
+      "epoch": 41.15,
+      "learning_rate": 1.4113740458015266e-05,
+      "loss": 1.1595,
+      "step": 28600
+    },
+    {
+      "epoch": 41.29,
+      "learning_rate": 1.388473282442748e-05,
+      "loss": 1.1527,
+      "step": 28700
+    },
+    {
+      "epoch": 41.44,
+      "learning_rate": 1.3655725190839693e-05,
+      "loss": 1.1517,
+      "step": 28800
+    },
+    {
+      "epoch": 41.58,
+      "learning_rate": 1.3426717557251907e-05,
+      "loss": 1.1609,
+      "step": 28900
+    },
+    {
+      "epoch": 41.73,
+      "learning_rate": 1.3197709923664122e-05,
+      "loss": 1.1668,
+      "step": 29000
+    },
+    {
+      "epoch": 41.73,
+      "eval_cer": 0.10246417886946937,
+      "eval_loss": 0.23187227547168732,
+      "eval_runtime": 40.5813,
+      "eval_samples_per_second": 11.237,
+      "eval_steps_per_second": 1.405,
+      "eval_wer": 0.3187858652974932,
+      "step": 29000
+    },
+    {
+      "epoch": 41.87,
+      "learning_rate": 1.2968702290076335e-05,
+      "loss": 1.1536,
+      "step": 29100
+    },
+    {
+      "epoch": 42.01,
+      "learning_rate": 1.2739694656488549e-05,
+      "loss": 1.1649,
+      "step": 29200
+    },
+    {
+      "epoch": 42.16,
+      "learning_rate": 1.2510687022900762e-05,
+      "loss": 1.1459,
+      "step": 29300
+    },
+    {
+      "epoch": 42.3,
+      "learning_rate": 1.2281679389312975e-05,
+      "loss": 1.1495,
+      "step": 29400
+    },
+    {
+      "epoch": 42.45,
+      "learning_rate": 1.205267175572519e-05,
+      "loss": 1.1448,
+      "step": 29500
+    },
+    {
+      "epoch": 42.45,
+      "eval_cer": 0.09947252401196661,
+      "eval_loss": 0.2292834371328354,
+      "eval_runtime": 41.732,
+      "eval_samples_per_second": 10.927,
+      "eval_steps_per_second": 1.366,
+      "eval_wer": 0.3098761703412866,
+      "step": 29500
+    },
+    {
+      "epoch": 42.59,
+      "learning_rate": 1.1823664122137404e-05,
+      "loss": 1.1408,
+      "step": 29600
+    },
+    {
+      "epoch": 42.73,
+      "learning_rate": 1.1594656488549618e-05,
+      "loss": 1.1458,
+      "step": 29700
+    },
+    {
+      "epoch": 42.88,
+      "learning_rate": 1.1365648854961831e-05,
+      "loss": 1.1358,
+      "step": 29800
+    },
+    {
+      "epoch": 43.02,
+      "learning_rate": 1.1136641221374044e-05,
+      "loss": 1.1519,
+      "step": 29900
+    },
+    {
+      "epoch": 43.17,
+      "learning_rate": 1.0909923664122137e-05,
+      "loss": 1.1327,
+      "step": 30000
+    },
+    {
+      "epoch": 43.17,
+      "eval_cer": 0.09793733270351125,
+      "eval_loss": 0.2265164852142334,
+      "eval_runtime": 40.9338,
+      "eval_samples_per_second": 11.14,
+      "eval_steps_per_second": 1.392,
+      "eval_wer": 0.3047417698580489,
+      "step": 30000
+    },
+    {
+      "epoch": 43.31,
+      "learning_rate": 1.068091603053435e-05,
+      "loss": 1.1322,
+      "step": 30100
+    },
+    {
+      "epoch": 43.45,
+      "learning_rate": 1.0451908396946564e-05,
+      "loss": 1.1392,
+      "step": 30200
+    },
+    {
+      "epoch": 43.6,
+      "learning_rate": 1.0222900763358777e-05,
+      "loss": 1.1318,
+      "step": 30300
+    },
+    {
+      "epoch": 43.74,
+      "learning_rate": 9.99389312977099e-06,
+      "loss": 1.1321,
+      "step": 30400
+    },
+    {
+      "epoch": 43.88,
+      "learning_rate": 9.764885496183206e-06,
+      "loss": 1.1307,
+      "step": 30500
+    },
+    {
+      "epoch": 43.88,
+      "eval_cer": 0.09888206581640686,
+      "eval_loss": 0.22221311926841736,
+      "eval_runtime": 40.9097,
+      "eval_samples_per_second": 11.147,
+      "eval_steps_per_second": 1.393,
+      "eval_wer": 0.30776200543642407,
+      "step": 30500
+    },
+    {
+      "epoch": 44.03,
+      "learning_rate": 9.53587786259542e-06,
+      "loss": 1.1358,
+      "step": 30600
+    },
+    {
+      "epoch": 44.17,
+      "learning_rate": 9.306870229007633e-06,
+      "loss": 1.1342,
+      "step": 30700
+    },
+    {
+      "epoch": 44.32,
+      "learning_rate": 9.077862595419846e-06,
+      "loss": 1.1348,
+      "step": 30800
+    },
+    {
+      "epoch": 44.46,
+      "learning_rate": 8.84885496183206e-06,
+      "loss": 1.1294,
+      "step": 30900
+    },
+    {
+      "epoch": 44.6,
+      "learning_rate": 8.619847328244275e-06,
+      "loss": 1.1419,
+      "step": 31000
+    },
+    {
+      "epoch": 44.6,
+      "eval_cer": 0.09813415210203118,
+      "eval_loss": 0.22149430215358734,
+      "eval_runtime": 40.8027,
+      "eval_samples_per_second": 11.176,
+      "eval_steps_per_second": 1.397,
+      "eval_wer": 0.3038356991845364,
+      "step": 31000
+    },
+    {
+      "epoch": 44.75,
+      "learning_rate": 8.390839694656488e-06,
+      "loss": 1.1191,
+      "step": 31100
+    },
+    {
+      "epoch": 44.89,
+      "learning_rate": 8.161832061068702e-06,
+      "loss": 1.1223,
+      "step": 31200
+    },
+    {
+      "epoch": 45.04,
+      "learning_rate": 7.932824427480915e-06,
+      "loss": 1.1393,
+      "step": 31300
+    },
+    {
+      "epoch": 45.18,
+      "learning_rate": 7.703816793893129e-06,
+      "loss": 1.1172,
+      "step": 31400
+    },
+    {
+      "epoch": 45.32,
+      "learning_rate": 7.474809160305343e-06,
+      "loss": 1.1231,
+      "step": 31500
+    },
+    {
+      "epoch": 45.32,
+      "eval_cer": 0.09722878286883956,
+      "eval_loss": 0.2193477302789688,
+      "eval_runtime": 40.6396,
+      "eval_samples_per_second": 11.221,
+      "eval_steps_per_second": 1.403,
+      "eval_wer": 0.3012684989429176,
+      "step": 31500
+    },
+    {
+      "epoch": 45.47,
+      "learning_rate": 7.245801526717557e-06,
+      "loss": 1.1289,
+      "step": 31600
+    },
+    {
+      "epoch": 45.61,
+      "learning_rate": 7.016793893129771e-06,
+      "loss": 1.1083,
+      "step": 31700
+    },
+    {
+      "epoch": 45.75,
+      "learning_rate": 6.787786259541984e-06,
+      "loss": 1.109,
+      "step": 31800
+    },
+    {
+      "epoch": 45.9,
+      "learning_rate": 6.558778625954198e-06,
+      "loss": 1.1218,
+      "step": 31900
+    },
+    {
+      "epoch": 46.04,
+      "learning_rate": 6.329770992366412e-06,
+      "loss": 1.139,
+      "step": 32000
+    },
+    {
+      "epoch": 46.04,
+      "eval_cer": 0.09683514407179972,
+      "eval_loss": 0.2162453532218933,
+      "eval_runtime": 41.1951,
+      "eval_samples_per_second": 11.069,
+      "eval_steps_per_second": 1.384,
+      "eval_wer": 0.30066445182724255,
+      "step": 32000
+    },
+    {
+      "epoch": 46.19,
+      "learning_rate": 6.100763358778626e-06,
+      "loss": 1.1155,
+      "step": 32100
+    },
+    {
+      "epoch": 46.33,
+      "learning_rate": 5.8717557251908395e-06,
+      "loss": 1.1212,
+      "step": 32200
+    },
+    {
+      "epoch": 46.47,
+      "learning_rate": 5.642748091603053e-06,
+      "loss": 1.1149,
+      "step": 32300
+    },
+    {
+      "epoch": 46.62,
+      "learning_rate": 5.413740458015267e-06,
+      "loss": 1.1183,
+      "step": 32400
+    },
+    {
+      "epoch": 46.76,
+      "learning_rate": 5.184732824427481e-06,
+      "loss": 1.1114,
+      "step": 32500
+    },
+    {
+      "epoch": 46.76,
+      "eval_cer": 0.09596913871831207,
+      "eval_loss": 0.2121613770723343,
+      "eval_runtime": 40.6982,
+      "eval_samples_per_second": 11.204,
+      "eval_steps_per_second": 1.401,
+      "eval_wer": 0.2982482633645424,
+      "step": 32500
+    },
+    {
+      "epoch": 46.91,
+      "learning_rate": 4.955725190839695e-06,
+      "loss": 1.1091,
+      "step": 32600
+    },
+    {
+      "epoch": 47.05,
+      "learning_rate": 4.726717557251908e-06,
+      "loss": 1.1148,
+      "step": 32700
+    },
+    {
+      "epoch": 47.19,
+      "learning_rate": 4.497709923664122e-06,
+      "loss": 1.0962,
+      "step": 32800
+    },
+    {
+      "epoch": 47.34,
+      "learning_rate": 4.268702290076335e-06,
+      "loss": 1.0984,
+      "step": 32900
+    },
+    {
+      "epoch": 47.48,
+      "learning_rate": 4.03969465648855e-06,
+      "loss": 1.111,
+      "step": 33000
+    },
+    {
+      "epoch": 47.48,
+      "eval_cer": 0.09482758620689655,
+      "eval_loss": 0.21248506009578705,
+      "eval_runtime": 40.6368,
+      "eval_samples_per_second": 11.221,
+      "eval_steps_per_second": 1.403,
+      "eval_wer": 0.2946239806704923,
+      "step": 33000
+    },
+    {
+      "epoch": 47.63,
+      "learning_rate": 3.810687022900763e-06,
+      "loss": 1.1031,
+      "step": 33100
+    },
+    {
+      "epoch": 47.77,
+      "learning_rate": 3.581679389312977e-06,
+      "loss": 1.1159,
+      "step": 33200
+    },
+    {
+      "epoch": 47.91,
+      "learning_rate": 3.352671755725191e-06,
+      "loss": 1.0905,
+      "step": 33300
+    },
+    {
+      "epoch": 48.06,
+      "learning_rate": 3.1236641221374048e-06,
+      "loss": 1.1087,
+      "step": 33400
+    },
+    {
+      "epoch": 48.2,
+      "learning_rate": 2.894656488549618e-06,
+      "loss": 1.0982,
+      "step": 33500
+    },
+    {
+      "epoch": 48.2,
+      "eval_cer": 0.09533931664304834,
+      "eval_loss": 0.2098563313484192,
+      "eval_runtime": 40.5946,
+      "eval_samples_per_second": 11.233,
+      "eval_steps_per_second": 1.404,
+      "eval_wer": 0.2956810631229236,
+      "step": 33500
+    },
+    {
+      "epoch": 48.34,
+      "learning_rate": 2.66793893129771e-06,
+      "loss": 1.0947,
+      "step": 33600
+    },
+    {
+      "epoch": 48.49,
+      "learning_rate": 2.4389312977099237e-06,
+      "loss": 1.1102,
+      "step": 33700
+    },
+    {
+      "epoch": 48.63,
+      "learning_rate": 2.209923664122137e-06,
+      "loss": 1.0891,
+      "step": 33800
+    },
+    {
+      "epoch": 48.78,
+      "learning_rate": 1.980916030534351e-06,
+      "loss": 1.0937,
+      "step": 33900
+    },
+    {
+      "epoch": 48.92,
+      "learning_rate": 1.7519083969465647e-06,
+      "loss": 1.109,
+      "step": 34000
+    },
+    {
+      "epoch": 48.92,
+      "eval_cer": 0.09545740828216029,
+      "eval_loss": 0.20918497443199158,
+      "eval_runtime": 40.877,
+      "eval_samples_per_second": 11.155,
+      "eval_steps_per_second": 1.394,
+      "eval_wer": 0.29553005134400484,
+      "step": 34000
+    },
+    {
+      "epoch": 49.06,
+      "learning_rate": 1.5229007633587786e-06,
+      "loss": 1.097,
+      "step": 34100
+    },
+    {
+      "epoch": 49.21,
+      "learning_rate": 1.2938931297709922e-06,
+      "loss": 1.0909,
+      "step": 34200
+    },
+    {
+      "epoch": 49.35,
+      "learning_rate": 1.0648854961832059e-06,
+      "loss": 1.1008,
+      "step": 34300
+    },
+    {
+      "epoch": 49.5,
+      "learning_rate": 8.358778625954198e-07,
+      "loss": 1.0904,
+      "step": 34400
+    },
+    {
+      "epoch": 49.64,
+      "learning_rate": 6.068702290076335e-07,
+      "loss": 1.0905,
+      "step": 34500
+    },
+    {
+      "epoch": 49.64,
+      "eval_cer": 0.09526058888364038,
+      "eval_loss": 0.20883652567863464,
+      "eval_runtime": 40.602,
+      "eval_samples_per_second": 11.231,
+      "eval_steps_per_second": 1.404,
+      "eval_wer": 0.2953790395650861,
+      "step": 34500
+    },
+    {
+      "epoch": 49.78,
+      "learning_rate": 3.778625954198473e-07,
+      "loss": 1.0961,
+      "step": 34600
+    },
+    {
+      "epoch": 49.93,
+      "learning_rate": 1.4885496183206107e-07,
+      "loss": 1.095,
+      "step": 34700
+    },
+    {
+      "epoch": 50.0,
+      "step": 34750,
+      "total_flos": 2.8392187465644065e+20,
+      "train_loss": 2.2316733406121783,
+      "train_runtime": 114311.9751,
+      "train_samples_per_second": 9.737,
+      "train_steps_per_second": 0.304
+    }
+  ],
+  "max_steps": 34750,
+  "num_train_epochs": 50,
+  "total_flos": 2.8392187465644065e+20,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4af57085f0712099c06c191a8e3123d5fbba4119a615b1c3ec5ef78e066139b2
+size 2991

vocab.json

@@ -0,0 +1 @@
+{"가": 1, "각": 2, "간": 3, "갈": 4, "감": 5, "갑": 6, "값": 7, "갓": 8, "갔": 9, "강": 10, "갖": 11, "같": 12, "갚": 13, "개": 14, "객": 15, "갠": 16, "갤": 17, "갯": 18, "갱": 19, "거": 20, "걱": 21, "건": 22, "걷": 23, "걸": 24, "검": 25, "겁": 26, "것": 27, "겉": 28, "게": 29, "겐": 30, "겟": 31, "겠": 32, "겨": 33, "격": 34, "겪": 35, "견": 36, "결": 37, "겸": 38, "겹": 39, "겼": 40, "경": 41, "곁": 42, "계": 43, "고": 44, "곡": 45, "곤": 46, "곧": 47, "골": 48, "곰": 49, "곱": 50, "곳": 51, "공": 52, "곶": 53, "과": 54, "곽": 55, "관": 56, "괄": 57, "괌": 58, "광": 59, "괘": 60, "괭": 61, "괴": 62, "굉": 63, "교": 64, "구": 65, "국": 66, "군": 67, "굳": 68, "굴": 69, "굵": 70, "굶": 71, "굽": 72, "굿": 73, "궁": 74, "궈": 75, "권": 76, "궐": 77, "궤": 78, "귀": 79, "규": 80, "균": 81, "그": 82, "극": 83, "근": 84, "글": 85, "금": 86, "급": 87, "긋": 88, "긍": 89, "기": 90, "긴": 91, "길": 92, "김": 93, "깁": 94, "깃": 95, "깊": 96, "까": 97, "깎": 98, "깔": 99, "깝": 100, "깡": 101, "깥": 102, "깨": 103, "깬": 104, "꺼": 105, "껍": 106, "껏": 107, "껑": 108, "께": 109, "껴": 110, "꼈": 111, "꼬": 112, "꼭": 113, "꼴": 114, "꼼": 115, "꼽": 116, "꽁": 117, "꽂": 118, "꽃": 119, "꽝": 120, "꽥": 121, "꾸": 122, "꾼": 123, "꿀": 124, "꿇": 125, "꿈": 126, "꿎": 127, "꿔": 128, "꿨": 129, "꿰": 130, "뀌": 131, "뀐": 132, "끄": 133, "끅": 134, "끈": 135, "끊": 136, "끌": 137, "끓": 138, "끔": 139, "끗": 140, "끝": 141, "끼": 142, "끽": 143, "낀": 144, "낌": 145, "나": 146, "낙": 147, "난": 148, "날": 149, "낡": 150, "남": 151, "납": 152, "낫": 153, "났": 154, "낭": 155, "낮": 156, "낯": 157, "낳": 158, "내": 159, "낸": 160, "낼": 161, "냄": 162, "냅": 163, "냇": 164, "냈": 165, "냉": 166, "냐": 167, "냥": 168, "너": 169, "넉": 170, "넌": 171, "널": 172, "넓": 173, "넘": 174, "넛": 175, "넣": 176, "네": 177, "넥": 178, "넨": 179, "넷": 180, "녀": 181, "녁": 182, "년": 183, "념": 184, "녔": 185, "녕": 186, "녘": 187, "노": 188, "녹": 189, "논": 190, "놀": 191, "놈": 192, "농": 193, "높": 194, "놓": 195, "놨": 196, "뇌": 197, "뇨": 198, "뇽": 199, "누": 200, "눈": 201, "눌": 202, "눠": 203, "뉜": 204, "뉴": 205, "늉": 206, "느": 207, "는": 208, "늘": 209, "늠": 210, "능": 211, "늦": 212, "늪": 213, "늬": 214, "니": 215, "닉": 216, "닌": 217, "닐": 218, "님": 219, "닙": 220, "닛": 221, "닝": 222, "다": 223, "닥": 224, "닦": 225, "단": 226, "닫": 227, "달": 228, "닭": 229, "닮": 230, "닳": 231, "담": 232, "답": 233, "닷": 234, "당": 235, "닿": 236, "대": 237, "댁": 238, "댐": 239, "댓": 240, "더": 241, "덕": 242, "던": 243, "덜": 244, "덟": 245, "덤": 246, "덧": 247, "덩": 248, "덫": 249, "덮": 250, "데": 251, "덴": 252, "델": 253, "뎅": 254, "뎌": 255, "도": 256, "독": 257, "돈": 258, "돋": 259, "돌": 260, "돔": 261, "돕": 262, "돗": 263, "동": 264, "돛": 265, "돼": 266, "됐": 267, "되": 268, "된": 269, "될": 270, "됨": 271, "됩": 272, "두": 273, "둑": 274, "둔": 275, "둘": 276, "둠": 277, "둥": 278, "둬": 279, "뒀": 280, "뒤": 281, "뒷": 282, "듀": 283, "드": 284, "득": 285, "든": 286, "듣": 287, "들": 288, "듬": 289, "듭": 290, "듯": 291, "등": 292, "디": 293, "딘": 294, "딜": 295, "딥": 296, "딨": 297, "딩": 298, "딪": 299, "따": 300, "딱": 301, "딴": 302, "딸": 303, "땀": 304, "땄": 305, "땅": 306, "때": 307, "땐": 308, "땠": 309, "땡": 310, "떠": 311, "떡": 312, "떤": 313, "떨": 314, "떴": 315, "떻": 316, "떼": 317, "뗄": 318, "또": 319, "똑": 320, "똥": 321, "뚜": 322, "뚝": 323, "뚫": 324, "뛰": 325, "뛴": 326, "뜨": 327, "뜯": 328, "뜸": 329, "뜻": 330, "띄": 331, "띔": 332, "띠": 333, "띤": 334, "띨": 335, "띵": 336, "라": 337, "락": 338, "란": 339, "랄": 340, "람": 341, "랍": 342, "랐": 343, "랑": 344, "래": 345, "랙": 346, "랜": 347, "램": 348, "랫": 349, "랬": 350, "랭": 351, "랴": 352, "략": 353, "량": 354, "러": 355, "럭": 356, "런": 357, "럴": 358, "럼": 359, "럽": 360, "렀": 361, "렁": 362, "렇": 363, "레": 364, "렉": 365, "렌": 366, "렘": 367, "렛": 368, "려": 369, "력": 370, "련": 371, "렬": 372, "렴": 373, "렵": 374, "렷": 375, "렸": 376, "령": 377, "례": 378, "로": 379, "록": 380, "론": 381, "롤": 382, "롬": 383, "롭": 384, "롯": 385, "롱": 386, "뢰": 387, "료": 388, "룡": 389, "루": 390, "룬": 391, "룰": 392, "룸": 393, "룹": 394, "뤄": 395, "뤘": 396, "뤼": 397, "류": 398, "륙": 399, "륜": 400, "률": 401, "륭": 402, "르": 403, "륵": 404, "른": 405, "를": 406, "름": 407, "릅": 408, "릇": 409, "릉": 410, "릎": 411, "리": 412, "릭": 413, "린": 414, "릴": 415, "림": 416, "립": 417, "릿": 418, "링": 419, "마": 420, "막": 421, "만": 422, "많": 423, "맏": 424, "말": 425, "맑": 426, "맘": 427, "맙": 428, "맛": 429, "망": 430, "맞": 431, "맡": 432, "매": 433, "맥": 434, "맨": 435, "맹": 436, "맺": 437, "머": 438, "먹": 439, "먼": 440, "멀": 441, "멈": 442, "멋": 443, "멍": 444, "메": 445, "멕": 446, "멘": 447, "멜": 448, "멧": 449, "며": 450, "면": 451, "멸": 452, "명": 453, "몇": 454, "모": 455, "목": 456, "몫": 457, "몬": 458, "몰": 459, "몸": 460, "못": 461, "몽": 462, "묘": 463, "무": 464, "묵": 465, "묶": 466, "문": 467, "묻": 468, "물": 469, "뭄": 470, "뭇": 471, "뭐": 472, "뭔": 473, "뭘": 474, "뮤": 475, "뮬": 476, "므": 477, "믈": 478, "미": 479, "믹": 480, "민": 481, "믿": 482, "밀": 483, "밋": 484, "밌": 485, "밍": 486, "및": 487, "밑": 488, "바": 489, "박": 490, "밖": 491, "반": 492, "받": 493, "발": 494, "밝": 495, "밤": 496, "밥": 497, "방": 498, "밭": 499, "배": 500, "백": 501, "밴": 502, "뱀": 503, "뱃": 504, "뱅": 505, "버": 506, "벅": 507, "번": 508, "벌": 509, "범": 510, "법": 511, "벗": 512, "벚": 513, "베": 514, "벤": 515, "벨": 516, "벳": 517, "벼": 518, "벽": 519, "변": 520, "별": 521, "볍": 522, "볐": 523, "병": 524, "볕": 525, "보": 526, "복": 527, "볶": 528, "본": 529, "볼": 530, "봄": 531, "봅": 532, "봇": 533, "봉": 534, "봐": 535, "봤": 536, "뵈": 537, "뵙": 538, "부": 539, "북": 540, "분": 541, "불": 542, "붉": 543, "붐": 544, "붓": 545, "붕": 546, "붙": 547, "뷔": 548, "뷰": 549, "브": 550, "븐": 551, "블": 552, "비": 553, "빅": 554, "빈": 555, "빌": 556, "빗": 557, "빙": 558, "빚": 559, "빛": 560, "빠": 561, "빨": 562, "빵": 563, "빼": 564, "뺀": 565, "뺌": 566, "뺏": 567, "뺑": 568, "뻐": 569, "뻑": 570, "뻔": 571, "뻗": 572, "뻘": 573, "뼈": 574, "뽀": 575, "뽑": 576, "뽕": 577, "뿌": 578, "뿐": 579, "뿜": 580, "쁘": 581, "쁜": 582, "쁩": 583, "삐": 584, "사": 585, "삭": 586, "산": 587, "살": 588, "삶": 589, "삼": 590, "삽": 591, "삿": 592, "샀": 593, "상": 594, "새": 595, "색": 596, "샌": 597, "샐": 598, "샘": 599, "생": 600, "샤": 601, "샬": 602, "샵": 603, "샷": 604, "서": 605, "석": 606, "섞": 607, "선": 608, "섣": 609, "설": 610, "섬": 611, "섭": 612, "섯": 613, "섰": 614, "성": 615, "세": 616, "섹": 617, "센": 618, "셀": 619, "셈": 620, "셉": 621, "셋": 622, "셔": 623, "션": 624, "셜": 625, "셨": 626, "셰": 627, "소": 628, "속": 629, "손": 630, "솔": 631, "솜": 632, "솟": 633, "송": 634, "솥": 635, "쇄": 636, "쇠": 637, "쇤": 638, "쇼": 639, "숍": 640, "수": 641, "숙": 642, "순": 643, "술": 644, "숨": 645, "숭": 646, "숲": 647, "쉬": 648, "쉰": 649, "쉼": 650, "쉽": 651, "슈": 652, "슐": 653, "스": 654, "슨": 655, "슬": 656, "슴": 657, "습": 658, "슷": 659, "승": 660, "시": 661, "식": 662, "신": 663, "실": 664, "싫": 665, "심": 666, "십": 667, "싱": 668, "싶": 669, "싸": 670, "싹": 671, "싼": 672, "쌀": 673, "쌈": 674, "쌌": 675, "쌍": 676, "쌓": 677, "써": 678, "썩": 679, "썰": 680, "썼": 681, "쏘": 682, "쏜": 683, "쏟": 684, "쏠": 685, "쑤": 686, "쓰": 687, "쓴": 688, "쓸": 689, "씀": 690, "씁": 691, "씌": 692, "씨": 693, "씩": 694, "씬": 695, "씸": 696, "씻": 697, "아": 698, "악": 699, "안": 700, "앉": 701, "않": 702, "알": 703, "앓": 704, "암": 705, "압": 706, "앗": 707, "았": 708, "앙": 709, "앞": 710, "애": 711, "액": 712, "앤": 713, "앨": 714, "앱": 715, "앵": 716, "야": 717, "약": 718, "얇": 719, "양": 720, "얗": 721, "얘": 722, "어": 723, "억": 724, "언": 725, "얹": 726, "얻": 727, "얼": 728, "얽": 729, "엄": 730, "업": 731, "없": 732, "엇": 733, "었": 734, "엉": 735, "엎": 736, "에": 737, "엑": 738, "엔": 739, "엘": 740, "엠": 741, "엡": 742, "엣": 743, "여": 744, "역": 745, "연": 746, "열": 747, "엷": 748, "염": 749, "엽": 750, "엿": 751, "였": 752, "영": 753, "옆": 754, "예": 755, "옛": 756, "오": 757, "옥": 758, "온": 759, "올": 760, "옮": 761, "옳": 762, "옴": 763, "옵": 764, "옷": 765, "옹": 766, "와": 767, "완": 768, "왈": 769, "왔": 770, "왕": 771, "왜": 772, "외": 773, "왼": 774, "요": 775, "욕": 776, "용": 777, "우": 778, "욱": 779, "운": 780, "울": 781, "움": 782, "웁": 783, "웃": 784, "웅": 785, "워": 786, "원": 787, "월": 788, "웠": 789, "웨": 790, "웬": 791, "웹": 792, "위": 793, "윈": 794, "윌": 795, "윗": 796, "윙": 797, "유": 798, "육": 799, "윤": 800, "율": 801, "융": 802, "으": 803, "은": 804, "을": 805, "음": 806, "읍": 807, "응": 808, "의": 809, "이": 810, "익": 811, "인": 812, "일": 813, "읽": 814, "잃": 815, "임": 816, "입": 817, "잇": 818, "있": 819, "잉": 820, "잊": 821, "잎": 822, "자": 823, "작": 824, "잔": 825, "잖": 826, "잘": 827, "잠": 828, "잡": 829, "잣": 830, "장": 831, "잦": 832, "재": 833, "잭": 834, "잰": 835, "잽": 836, "쟁": 837, "저": 838, "적": 839, "전": 840, "절": 841, "젊": 842, "젋": 843, "점": 844, "접": 845, "젓": 846, "정": 847, "젖": 848, "제": 849, "젝": 850, "젠": 851, "젤": 852, "져": 853, "젼": 854, "졌": 855, "조": 856, "족": 857, "존": 858, "졸": 859, "좀": 860, "좁": 861, "종": 862, "좋": 863, "좌": 864, "죄": 865, "죠": 866, "주": 867, "죽": 868, "준": 869, "줄": 870, "줍": 871, "중": 872, "줘": 873, "줬": 874, "쥐": 875, "쥔": 876, "쥘": 877, "쥬": 878, "즈": 879, "즉": 880, "즌": 881, "즐": 882, "즘": 883, "증": 884, "지": 885, "직": 886, "진": 887, "질": 888, "짊": 889, "짐": 890, "집": 891, "짓": 892, "징": 893, "짖": 894, "짙": 895, "짚": 896, "짜": 897, "짝": 898, "짧": 899, "짬": 900, "째": 901, "쨌": 902, "쩌": 903, "쩍": 904, "쩔": 905, "쩜": 906, "쪼": 907, "쪽": 908, "쫄": 909, "쫓": 910, "쭉": 911, "쯤": 912, "찌": 913, "찍": 914, "찔": 915, "찢": 916, "찧": 917, "차": 918, "착": 919, "찬": 920, "찮": 921, "찰": 922, "참": 923, "찻": 924, "창": 925, "찾": 926, "채": 927, "책": 928, "챌": 929, "챔": 930, "챙": 931, "챠": 932, "처": 933, "척": 934, "천": 935, "철": 936, "첨": 937, "첩": 938, "첫": 939, "청": 940, "체": 941, "첸": 942, "첼": 943, "쳇": 944, "쳐": 945, "쳤": 946, "초": 947, "촉": 948, "촌": 949, "촘": 950, "촛": 951, "총": 952, "촨": 953, "촬": 954, "최": 955, "추": 956, "축": 957, "춘": 958, "출": 959, "춤": 960, "충": 961, "춰": 962, "취": 963, "츠": 964, "측": 965, "츰": 966, "층": 967, "치": 968, "칙": 969, "친": 970, "칠": 971, "침": 972, "칩": 973, "칫": 974, "칭": 975, "카": 976, "칸": 977, "칼": 978, "캉": 979, "캐": 980, "캔": 981, "캘": 982, "캠": 983, "커": 984, "컥": 985, "컨": 986, "컫": 987, "컴": 988, "컵": 989, "컷": 990, "컸": 991, "케": 992, "켈": 993, "켐": 994, "켑": 995, "켓": 996, "켜": 997, "켰": 998, "코": 999, "콘": 1000, "콜": 1001, "콤": 1002, "콥": 1003, "콧": 1004, "콩": 1005, "쾌": 1006, "쿄": 1007, "쿠": 1008, "쿡": 1009, "쿨": 1010, "쿼": 1011, "퀴": 1012, "큐": 1013, "크": 1014, "큰": 1015, "클": 1016, "큼": 1017, "키": 1018, "킥": 1019, "킨": 1020, "킬": 1021, "킷": 1022, "킹": 1023, "타": 1024, "탁": 1025, "탄": 1026, "탈": 1027, "탐": 1028, "탑": 1029, "탓": 1030, "탕": 1031, "태": 1032, "택": 1033, "탠": 1034, "탬": 1035, "탱": 1036, "터": 1037, "턱": 1038, "턴": 1039, "털": 1040, "텃": 1041, "텅": 1042, "테": 1043, "텍": 1044, "텐": 1045, "텔": 1046, "템": 1047, "텼": 1048, "토": 1049, "톡": 1050, "톤": 1051, "톨": 1052, "톰": 1053, "통": 1054, "퇴": 1055, "투": 1056, "툴": 1057, "툼": 1058, "퉁": 1059, "튀": 1060, "튜": 1061, "튬": 1062, "트": 1063, "특": 1064, "튼": 1065, "튿": 1066, "틀": 1067, "틈": 1068, "티": 1069, "틱": 1070, "틴": 1071, "틸": 1072, "팀": 1073, "팅": 1074, "파": 1075, "팍": 1076, "팎": 1077, "판": 1078, "팔": 1079, "팜": 1080, "팡": 1081, "패": 1082, "팩": 1083, "팬": 1084, "팰": 1085, "팻": 1086, "팽": 1087, "퍼": 1088, "펀": 1089, "펄": 1090, "펌": 1091, "페": 1092, "펜": 1093, "펠": 1094, "펫": 1095, "펴": 1096, "편": 1097, "펼": 1098, "폄": 1099, "폈": 1100, "평": 1101, "폐": 1102, "포": 1103, "폭": 1104, "폰": 1105, "폴": 1106, "폼": 1107, "표": 1108, "푸": 1109, "푹": 1110, "푼": 1111, "풀": 1112, "품": 1113, "풋": 1114, "풍": 1115, "퓨": 1116, "퓰": 1117, "프": 1118, "픈": 1119, "플": 1120, "픔": 1121, "피": 1122, "픽": 1123, "핀": 1124, "필": 1125, "핏": 1126, "핑": 1127, "하": 1128, "학": 1129, "한": 1130, "할": 1131, "함": 1132, "합": 1133, "핫": 1134, "항": 1135, "해": 1136, "핵": 1137, "핸": 1138, "햇": 1139, "했": 1140, "행": 1141, "향": 1142, "허": 1143, "헌": 1144, "헐": 1145, "험": 1146, "헝": 1147, "헤": 1148, "헨": 1149, "헬": 1150, "헴": 1151, "헷": 1152, "혀": 1153, "혁": 1154, "현": 1155, "혈": 1156, "혐": 1157, "협": 1158, "혔": 1159, "형": 1160, "혜": 1161, "호": 1162, "혹": 1163, "혼": 1164, "홀": 1165, "홈": 1166, "홉": 1167, "홍": 1168, "화": 1169, "확": 1170, "환": 1171, "활": 1172, "황": 1173, "회": 1174, "획": 1175, "횡": 1176, "효": 1177, "후": 1178, "훈": 1179, "훌": 1180, "훤": 1181, "훨": 1182, "훼": 1183, "휘": 1184, "휩": 1185, "휴": 1186, "흉": 1187, "흐": 1188, "흑": 1189, "흔": 1190, "흘": 1191, "흙": 1192, "흠": 1193, "흡": 1194, "흥": 1195, "흩": 1196, "희": 1197, "흰": 1198, "히": 1199, "힌": 1200, "힐": 1201, "힘": 1202, "|": 0, "[UNK]": 1203, "[PAD]": 1204}