add model files

.gitignore

@@ -0,0 +1,2 @@
+checkpoint-*/
+wandb/

README.md

@@ -1,3 +1,130 @@
----
-license: cc
----
+---
+language:
+- tr
+tags:
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_8_0
+- generated_from_trainer
+datasets:
+- common_voice
+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 [./checkpoint-1000](https://huggingface.co/./checkpoint-1000) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - TR dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.3282
+- Wer: 0.2836
+
+## 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: 0.0003
+- train_batch_size: 96
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 192
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 100
+- num_epochs: 100.0
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|
+| 1.0671        | 2.04  | 200   | 0.3079          | 0.2752 |
+| 0.6433        | 4.08  | 400   | 0.2728          | 0.2848 |
+| 0.5687        | 6.12  | 600   | 0.2882          | 0.3036 |
+| 0.5355        | 8.16  | 800   | 0.2778          | 0.2920 |
+| 0.5116        | 10.2  | 1000  | 0.2906          | 0.3014 |
+| 0.5313        | 9.16  | 1200  | 0.2984          | 0.3273 |
+| 0.4996        | 10.69 | 1400  | 0.3170          | 0.3344 |
+| 0.4845        | 12.21 | 1600  | 0.3202          | 0.3634 |
+| 0.5092        | 13.74 | 1800  | 0.3167          | 0.3373 |
+| 0.4777        | 15.27 | 2000  | 0.3292          | 0.3386 |
+| 0.4651        | 16.79 | 2200  | 0.3070          | 0.3427 |
+| 0.461         | 18.32 | 2400  | 0.3149          | 0.3561 |
+| 0.4481        | 19.85 | 2600  | 0.3292          | 0.3441 |
+| 0.4479        | 21.37 | 2800  | 0.3142          | 0.3209 |
+| 0.4305        | 22.9  | 3000  | 0.3525          | 0.3547 |
+| 0.4254        | 24.43 | 3200  | 0.3414          | 0.3400 |
+| 0.4066        | 25.95 | 3400  | 0.3118          | 0.3207 |
+| 0.4043        | 27.48 | 3600  | 0.3418          | 0.3483 |
+| 0.3985        | 29.01 | 3800  | 0.3254          | 0.3166 |
+| 0.3982        | 30.53 | 4000  | 0.3306          | 0.3453 |
+| 0.3929        | 32.06 | 4200  | 0.3262          | 0.3229 |
+| 0.378         | 33.59 | 4400  | 0.3546          | 0.3336 |
+| 0.4062        | 35.11 | 4600  | 0.3174          | 0.3457 |
+| 0.3648        | 36.64 | 4800  | 0.3377          | 0.3357 |
+| 0.3609        | 38.17 | 5000  | 0.3346          | 0.3520 |
+| 0.3483        | 39.69 | 5200  | 0.3350          | 0.3526 |
+| 0.3548        | 41.22 | 5400  | 0.3330          | 0.3406 |
+| 0.3446        | 42.75 | 5600  | 0.3398          | 0.3372 |
+| 0.3346        | 44.27 | 5800  | 0.3449          | 0.3288 |
+| 0.3309        | 45.8  | 6000  | 0.3320          | 0.3144 |
+| 0.326         | 47.33 | 6200  | 0.3400          | 0.3279 |
+| 0.3189        | 48.85 | 6400  | 0.3400          | 0.3150 |
+| 0.3165        | 50.38 | 6600  | 0.3359          | 0.2995 |
+| 0.3132        | 51.91 | 6800  | 0.3343          | 0.3096 |
+| 0.3092        | 53.44 | 7000  | 0.3224          | 0.3029 |
+| 0.2995        | 54.96 | 7200  | 0.3205          | 0.2985 |
+| 0.304         | 56.49 | 7400  | 0.3523          | 0.3034 |
+| 0.2952        | 58.02 | 7600  | 0.3289          | 0.2934 |
+| 0.2875        | 59.54 | 7800  | 0.3350          | 0.3008 |
+| 0.2868        | 61.07 | 8000  | 0.3537          | 0.3227 |
+| 0.2875        | 62.6  | 8200  | 0.3389          | 0.2970 |
+| 0.2778        | 64.12 | 8400  | 0.3370          | 0.2960 |
+| 0.2706        | 65.65 | 8600  | 0.3250          | 0.2802 |
+| 0.2669        | 67.18 | 8800  | 0.3351          | 0.2903 |
+| 0.2615        | 68.7  | 9000  | 0.3382          | 0.2989 |
+| 0.2563        | 70.23 | 9200  | 0.3312          | 0.2975 |
+| 0.2546        | 71.76 | 9400  | 0.3212          | 0.3003 |
+| 0.2482        | 73.28 | 9600  | 0.3337          | 0.3091 |
+| 0.2504        | 74.81 | 9800  | 0.3308          | 0.3110 |
+| 0.2456        | 76.34 | 10000 | 0.3157          | 0.3118 |
+| 0.2363        | 77.86 | 10200 | 0.3251          | 0.3144 |
+| 0.2319        | 79.39 | 10400 | 0.3253          | 0.3038 |
+| 0.2266        | 80.92 | 10600 | 0.3374          | 0.3038 |
+| 0.2279        | 82.44 | 10800 | 0.3268          | 0.2964 |
+| 0.2231        | 83.97 | 11000 | 0.3278          | 0.2950 |
+| 0.2185        | 85.5  | 11200 | 0.3462          | 0.2981 |
+| 0.2245        | 87.02 | 11400 | 0.3311          | 0.2895 |
+| 0.223         | 88.55 | 11600 | 0.3325          | 0.2877 |
+| 0.2121        | 90.08 | 11800 | 0.3337          | 0.2828 |
+| 0.2126        | 91.6  | 12000 | 0.3325          | 0.2808 |
+| 0.2027        | 93.13 | 12200 | 0.3277          | 0.2820 |
+| 0.2058        | 94.66 | 12400 | 0.3308          | 0.2827 |
+| 0.1991        | 96.18 | 12600 | 0.3279          | 0.2820 |
+| 0.1991        | 97.71 | 12800 | 0.3300          | 0.2822 |
+| 0.1986        | 99.24 | 13000 | 0.3285          | 0.2835 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.3
+- Tokenizers 0.11.0

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.32822880148887634,
+    "eval_runtime": 237.0221,
+    "eval_samples": 8339,
+    "eval_samples_per_second": 35.182,
+    "eval_steps_per_second": 4.4,
+    "eval_wer": 0.2835930339138405,
+    "train_loss": 0.29656382378731067,
+    "train_runtime": 73649.3567,
+    "train_samples": 25058,
+    "train_samples_per_second": 34.023,
+    "train_steps_per_second": 0.178
+}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "-", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e2", "\u00e7", "\u00eb", "\u00ee", "\u00f6", "\u00fc", "\u011f", "\u0131", "\u015f", "\u0307", "\u2047", "" ], "is_bpe": false}

config.json

@@ -0,0 +1,106 @@
+{
+  "_name_or_path": "./checkpoint-1000",
+  "activation_dropout": 0.055,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 256,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": false,
+  "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": true,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": false,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_norm": "group",
+  "feat_proj_dropout": 0.04,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.047,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.041,
+  "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": 0.4,
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 12,
+  "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": 12,
+  "num_negatives": 100,
+  "output_hidden_size": 768,
+  "pad_token_id": 39,
+  "proj_codevector_dim": 256,
+  "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": 40,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+    chars_to_ignore_regex = '[,?.!-;:""%\'"\'\'`…’»«‘“”�éû]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # note that order is important here!
+    token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    for t in token_sequences_to_ignore:
+        text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, data_dir=args.data_dir, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--data_dir", type=str, required=False, default=None,
+                        help="The directory contains the dataset")
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

eval_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.32822880148887634,
+    "eval_runtime": 237.0221,
+    "eval_samples": 8339,
+    "eval_samples_per_second": 35.182,
+    "eval_steps_per_second": 4.4,
+    "eval_wer": 0.2835930339138405
+}

language_model/5gram.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b750a609666264c2071c7fc4dec044155244557fab76815b229ffb6ee9172a0c
+size 1040566788

language_model/attrs.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5717124172f9f489cd469aa7632b2c999b8e577db0e9c8d585434d13d424400e
+size 377694615

run.sh

@@ -0,0 +1,41 @@
+export WANDB_ENTITY=cahya
+export WANDB_LOG_MODEL=true
+export WANDB_PROJECT=xlsr-turkish
+
+python run_speech_recognition_ctc.py \
+--dataset_name="mozilla-foundation/common_voice_8_0" \
+--model_name_or_path="./checkpoint-1000" \
+--dataset_config_name="tr" \
+--output_dir="./" \
+--overwrite_output_dir \
+--num_train_epochs="100" \
+--per_device_train_batch_size="96" \
+--per_device_eval_batch_size="8" \
+--gradient_accumulation_steps="2" \
+--learning_rate="3e-4" \
+--warmup_steps="100" \
+--length_column_name="input_length" \
+--evaluation_strategy="steps" \
+--text_column_name="sentence" \
+--save_steps="200" \
+--eval_steps="200" \
+--logging_steps="200" \
+--layerdrop="0.041" \
+--activation_dropout="0.055" \
+--attention_dropout="0.1" \
+--hidden_dropout="0.047" \
+--save_total_limit="3" \
+--freeze_feature_encoder \
+--feat_proj_dropout="0.04" \
+--mask_time_prob="0.4" \
+--mask_time_length="10" \
+--mask_feature_prob="0.25" \
+--mask_feature_length="64" \
+--gradient_checkpointing \
+--use_auth_token \
+--fp16=true \
+--group_by_length \
+--do_train=true \
+--do_eval=true \
+--push_to_hub=false \
+--chars_to_ignore , ? . ! \; \: \"\" \% \' \" \' \' \` … \’ » « \‘ '“' '”' � é û \( \)

run_speech_recognition_ctc.py

@@ -0,0 +1,748 @@
+#!/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+validation",
+        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"],
+        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]}
+
+    print(f"dataset: {datasets}")
+
+    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.
+    print("training_args.do_train:", training_args.do_train)
+    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()
+
+    print("do_train:", training_args.do_train)
+
+    if training_args.do_train:
+        print("load 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
+    )
+    print(f"char ignored: {data_args.chars_to_ignore} {chars_to_ignore_regex}")
+    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
+    )
+
+    print(f"config: {config}")
+
+    # 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,
+                )
+                print(f"vocab: {vocab_dict}")
+
+                # 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,
+        eos_token=None, bos_token=None,
+        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()

special_tokens_map.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "pad_token": "[PAD]"}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": null, "eos_token": null, "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": 100.0,
+    "train_loss": 0.29656382378731067,
+    "train_runtime": 73649.3567,
+    "train_samples": 25058,
+    "train_samples_per_second": 34.023,
+    "train_steps_per_second": 0.178
+}

trainer_state.json

@@ -0,0 +1,1000 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 100.0,
+  "global_step": 13100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 2.04,
+      "learning_rate": 0.0002969690721649484,
+      "loss": 1.0671,
+      "step": 200
+    },
+    {
+      "epoch": 2.04,
+      "eval_loss": 0.3079470694065094,
+      "eval_runtime": 237.7627,
+      "eval_samples_per_second": 35.073,
+      "eval_steps_per_second": 4.387,
+      "eval_wer": 0.2752062328139322,
+      "step": 200
+    },
+    {
+      "epoch": 4.08,
+      "learning_rate": 0.00029078350515463917,
+      "loss": 0.6433,
+      "step": 400
+    },
+    {
+      "epoch": 4.08,
+      "eval_loss": 0.27281925082206726,
+      "eval_runtime": 238.3892,
+      "eval_samples_per_second": 34.981,
+      "eval_steps_per_second": 4.375,
+      "eval_wer": 0.2847616865261228,
+      "step": 400
+    },
+    {
+      "epoch": 6.12,
+      "learning_rate": 0.0002845979381443299,
+      "loss": 0.5687,
+      "step": 600
+    },
+    {
+      "epoch": 6.12,
+      "eval_loss": 0.28817421197891235,
+      "eval_runtime": 238.0657,
+      "eval_samples_per_second": 35.028,
+      "eval_steps_per_second": 4.381,
+      "eval_wer": 0.3036205316223648,
+      "step": 600
+    },
+    {
+      "epoch": 8.16,
+      "learning_rate": 0.00027841237113402056,
+      "loss": 0.5355,
+      "step": 800
+    },
+    {
+      "epoch": 8.16,
+      "eval_loss": 0.27776163816452026,
+      "eval_runtime": 239.8353,
+      "eval_samples_per_second": 34.77,
+      "eval_steps_per_second": 4.349,
+      "eval_wer": 0.29200274977085244,
+      "step": 800
+    },
+    {
+      "epoch": 10.2,
+      "learning_rate": 0.00027222680412371134,
+      "loss": 0.5116,
+      "step": 1000
+    },
+    {
+      "epoch": 10.2,
+      "eval_loss": 0.2905969023704529,
+      "eval_runtime": 240.1935,
+      "eval_samples_per_second": 34.718,
+      "eval_steps_per_second": 4.342,
+      "eval_wer": 0.3013978001833181,
+      "step": 1000
+    },
+    {
+      "epoch": 9.16,
+      "learning_rate": 0.00027468461538461536,
+      "loss": 0.5313,
+      "step": 1200
+    },
+    {
+      "epoch": 9.16,
+      "eval_loss": 0.2984345555305481,
+      "eval_runtime": 234.3486,
+      "eval_samples_per_second": 35.584,
+      "eval_steps_per_second": 4.451,
+      "eval_wer": 0.327314390467461,
+      "step": 1200
+    },
+    {
+      "epoch": 10.69,
+      "learning_rate": 0.00027006923076923077,
+      "loss": 0.4996,
+      "step": 1400
+    },
+    {
+      "epoch": 10.69,
+      "eval_loss": 0.3169882297515869,
+      "eval_runtime": 237.8663,
+      "eval_samples_per_second": 35.058,
+      "eval_steps_per_second": 4.385,
+      "eval_wer": 0.3344179651695692,
+      "step": 1400
+    },
+    {
+      "epoch": 12.21,
+      "learning_rate": 0.0002654538461538461,
+      "loss": 0.4845,
+      "step": 1600
+    },
+    {
+      "epoch": 12.21,
+      "eval_loss": 0.32016345858573914,
+      "eval_runtime": 236.9291,
+      "eval_samples_per_second": 35.196,
+      "eval_steps_per_second": 4.402,
+      "eval_wer": 0.36338221814848765,
+      "step": 1600
+    },
+    {
+      "epoch": 13.74,
+      "learning_rate": 0.00026086153846153847,
+      "loss": 0.5092,
+      "step": 1800
+    },
+    {
+      "epoch": 13.74,
+      "eval_loss": 0.3166552186012268,
+      "eval_runtime": 236.2482,
+      "eval_samples_per_second": 35.298,
+      "eval_steps_per_second": 4.415,
+      "eval_wer": 0.3373052245646196,
+      "step": 1800
+    },
+    {
+      "epoch": 15.27,
+      "learning_rate": 0.0002562461538461538,
+      "loss": 0.4777,
+      "step": 2000
+    },
+    {
+      "epoch": 15.27,
+      "eval_loss": 0.32921522855758667,
+      "eval_runtime": 235.517,
+      "eval_samples_per_second": 35.407,
+      "eval_steps_per_second": 4.429,
+      "eval_wer": 0.3385655362053162,
+      "step": 2000
+    },
+    {
+      "epoch": 16.79,
+      "learning_rate": 0.0002516307692307692,
+      "loss": 0.4651,
+      "step": 2200
+    },
+    {
+      "epoch": 16.79,
+      "eval_loss": 0.3070097863674164,
+      "eval_runtime": 238.1366,
+      "eval_samples_per_second": 35.018,
+      "eval_steps_per_second": 4.38,
+      "eval_wer": 0.34271310724106324,
+      "step": 2200
+    },
+    {
+      "epoch": 18.32,
+      "learning_rate": 0.0002470153846153846,
+      "loss": 0.461,
+      "step": 2400
+    },
+    {
+      "epoch": 18.32,
+      "eval_loss": 0.3148922026157379,
+      "eval_runtime": 237.1783,
+      "eval_samples_per_second": 35.159,
+      "eval_steps_per_second": 4.398,
+      "eval_wer": 0.35609532538955085,
+      "step": 2400
+    },
+    {
+      "epoch": 19.85,
+      "learning_rate": 0.00024239999999999998,
+      "loss": 0.4481,
+      "step": 2600
+    },
+    {
+      "epoch": 19.85,
+      "eval_loss": 0.32919421792030334,
+      "eval_runtime": 236.0044,
+      "eval_samples_per_second": 35.334,
+      "eval_steps_per_second": 4.419,
+      "eval_wer": 0.34411090742438133,
+      "step": 2600
+    },
+    {
+      "epoch": 21.37,
+      "learning_rate": 0.00023778461538461536,
+      "loss": 0.4479,
+      "step": 2800
+    },
+    {
+      "epoch": 21.37,
+      "eval_loss": 0.3142247200012207,
+      "eval_runtime": 234.6363,
+      "eval_samples_per_second": 35.54,
+      "eval_steps_per_second": 4.445,
+      "eval_wer": 0.32089825847846015,
+      "step": 2800
+    },
+    {
+      "epoch": 22.9,
+      "learning_rate": 0.00023316923076923077,
+      "loss": 0.4305,
+      "step": 3000
+    },
+    {
+      "epoch": 22.9,
+      "eval_loss": 0.3525453507900238,
+      "eval_runtime": 236.821,
+      "eval_samples_per_second": 35.212,
+      "eval_steps_per_second": 4.404,
+      "eval_wer": 0.35467461044912924,
+      "step": 3000
+    },
+    {
+      "epoch": 24.43,
+      "learning_rate": 0.00022855384615384612,
+      "loss": 0.4254,
+      "step": 3200
+    },
+    {
+      "epoch": 24.43,
+      "eval_loss": 0.34136688709259033,
+      "eval_runtime": 235.4909,
+      "eval_samples_per_second": 35.411,
+      "eval_steps_per_second": 4.429,
+      "eval_wer": 0.3400091659028414,
+      "step": 3200
+    },
+    {
+      "epoch": 25.95,
+      "learning_rate": 0.00022393846153846153,
+      "loss": 0.4066,
+      "step": 3400
+    },
+    {
+      "epoch": 25.95,
+      "eval_loss": 0.3118491470813751,
+      "eval_runtime": 236.1268,
+      "eval_samples_per_second": 35.316,
+      "eval_steps_per_second": 4.417,
+      "eval_wer": 0.3207378551787351,
+      "step": 3400
+    },
+    {
+      "epoch": 27.48,
+      "learning_rate": 0.0002193230769230769,
+      "loss": 0.4043,
+      "step": 3600
+    },
+    {
+      "epoch": 27.48,
+      "eval_loss": 0.34181562066078186,
+      "eval_runtime": 235.2934,
+      "eval_samples_per_second": 35.441,
+      "eval_steps_per_second": 4.433,
+      "eval_wer": 0.3482584784601283,
+      "step": 3600
+    },
+    {
+      "epoch": 29.01,
+      "learning_rate": 0.0002147076923076923,
+      "loss": 0.3985,
+      "step": 3800
+    },
+    {
+      "epoch": 29.01,
+      "eval_loss": 0.32544735074043274,
+      "eval_runtime": 236.4374,
+      "eval_samples_per_second": 35.269,
+      "eval_steps_per_second": 4.411,
+      "eval_wer": 0.31663611365719524,
+      "step": 3800
+    },
+    {
+      "epoch": 30.53,
+      "learning_rate": 0.00021009230769230766,
+      "loss": 0.3982,
+      "step": 4000
+    },
+    {
+      "epoch": 30.53,
+      "eval_loss": 0.33056947588920593,
+      "eval_runtime": 239.8848,
+      "eval_samples_per_second": 34.763,
+      "eval_steps_per_second": 4.348,
+      "eval_wer": 0.3452795600366636,
+      "step": 4000
+    },
+    {
+      "epoch": 32.06,
+      "learning_rate": 0.00020547692307692307,
+      "loss": 0.3929,
+      "step": 4200
+    },
+    {
+      "epoch": 32.06,
+      "eval_loss": 0.3262433409690857,
+      "eval_runtime": 238.1544,
+      "eval_samples_per_second": 35.015,
+      "eval_steps_per_second": 4.38,
+      "eval_wer": 0.3228689275893675,
+      "step": 4200
+    },
+    {
+      "epoch": 33.59,
+      "learning_rate": 0.00020086153846153845,
+      "loss": 0.378,
+      "step": 4400
+    },
+    {
+      "epoch": 33.59,
+      "eval_loss": 0.3545904755592346,
+      "eval_runtime": 234.4177,
+      "eval_samples_per_second": 35.573,
+      "eval_steps_per_second": 4.449,
+      "eval_wer": 0.33357011915673696,
+      "step": 4400
+    },
+    {
+      "epoch": 35.11,
+      "learning_rate": 0.00019624615384615385,
+      "loss": 0.4062,
+      "step": 4600
+    },
+    {
+      "epoch": 35.11,
+      "eval_loss": 0.3174082636833191,
+      "eval_runtime": 235.3658,
+      "eval_samples_per_second": 35.43,
+      "eval_steps_per_second": 4.431,
+      "eval_wer": 0.34566911090742436,
+      "step": 4600
+    },
+    {
+      "epoch": 36.64,
+      "learning_rate": 0.0001916307692307692,
+      "loss": 0.3648,
+      "step": 4800
+    },
+    {
+      "epoch": 36.64,
+      "eval_loss": 0.3376729190349579,
+      "eval_runtime": 239.2202,
+      "eval_samples_per_second": 34.859,
+      "eval_steps_per_second": 4.36,
+      "eval_wer": 0.33572410632447297,
+      "step": 4800
+    },
+    {
+      "epoch": 38.17,
+      "learning_rate": 0.0001870153846153846,
+      "loss": 0.3609,
+      "step": 5000
+    },
+    {
+      "epoch": 38.17,
+      "eval_loss": 0.33455467224121094,
+      "eval_runtime": 236.9279,
+      "eval_samples_per_second": 35.196,
+      "eval_steps_per_second": 4.402,
+      "eval_wer": 0.351993583868011,
+      "step": 5000
+    },
+    {
+      "epoch": 39.69,
+      "learning_rate": 0.0001824,
+      "loss": 0.3483,
+      "step": 5200
+    },
+    {
+      "epoch": 39.69,
+      "eval_loss": 0.3349843919277191,
+      "eval_runtime": 235.432,
+      "eval_samples_per_second": 35.42,
+      "eval_steps_per_second": 4.43,
+      "eval_wer": 0.3525893675527039,
+      "step": 5200
+    },
+    {
+      "epoch": 41.22,
+      "learning_rate": 0.0001777846153846154,
+      "loss": 0.3548,
+      "step": 5400
+    },
+    {
+      "epoch": 41.22,
+      "eval_loss": 0.33302220702171326,
+      "eval_runtime": 237.3922,
+      "eval_samples_per_second": 35.128,
+      "eval_steps_per_second": 4.394,
+      "eval_wer": 0.34062786434463793,
+      "step": 5400
+    },
+    {
+      "epoch": 42.75,
+      "learning_rate": 0.00017316923076923075,
+      "loss": 0.3446,
+      "step": 5600
+    },
+    {
+      "epoch": 42.75,
+      "eval_loss": 0.3398281931877136,
+      "eval_runtime": 236.2326,
+      "eval_samples_per_second": 35.3,
+      "eval_steps_per_second": 4.415,
+      "eval_wer": 0.3372135655362053,
+      "step": 5600
+    },
+    {
+      "epoch": 44.27,
+      "learning_rate": 0.00016855384615384615,
+      "loss": 0.3346,
+      "step": 5800
+    },
+    {
+      "epoch": 44.27,
+      "eval_loss": 0.34486979246139526,
+      "eval_runtime": 236.4721,
+      "eval_samples_per_second": 35.264,
+      "eval_steps_per_second": 4.411,
+      "eval_wer": 0.3287809349220898,
+      "step": 5800
+    },
+    {
+      "epoch": 45.8,
+      "learning_rate": 0.00016393846153846153,
+      "loss": 0.3309,
+      "step": 6000
+    },
+    {
+      "epoch": 45.8,
+      "eval_loss": 0.3319507837295532,
+      "eval_runtime": 236.6263,
+      "eval_samples_per_second": 35.241,
+      "eval_steps_per_second": 4.408,
+      "eval_wer": 0.31439046746104493,
+      "step": 6000
+    },
+    {
+      "epoch": 47.33,
+      "learning_rate": 0.0001593230769230769,
+      "loss": 0.326,
+      "step": 6200
+    },
+    {
+      "epoch": 47.33,
+      "eval_loss": 0.3399747312068939,
+      "eval_runtime": 236.0765,
+      "eval_samples_per_second": 35.323,
+      "eval_steps_per_second": 4.418,
+      "eval_wer": 0.32786434463794684,
+      "step": 6200
+    },
+    {
+      "epoch": 48.85,
+      "learning_rate": 0.0001547076923076923,
+      "loss": 0.3189,
+      "step": 6400
+    },
+    {
+      "epoch": 48.85,
+      "eval_loss": 0.3399554491043091,
+      "eval_runtime": 236.2028,
+      "eval_samples_per_second": 35.304,
+      "eval_steps_per_second": 4.416,
+      "eval_wer": 0.31500916590284145,
+      "step": 6400
+    },
+    {
+      "epoch": 50.38,
+      "learning_rate": 0.0001500923076923077,
+      "loss": 0.3165,
+      "step": 6600
+    },
+    {
+      "epoch": 50.38,
+      "eval_loss": 0.3359447121620178,
+      "eval_runtime": 235.0525,
+      "eval_samples_per_second": 35.477,
+      "eval_steps_per_second": 4.437,
+      "eval_wer": 0.29945004582951423,
+      "step": 6600
+    },
+    {
+      "epoch": 51.91,
+      "learning_rate": 0.00014547692307692305,
+      "loss": 0.3132,
+      "step": 6800
+    },
+    {
+      "epoch": 51.91,
+      "eval_loss": 0.3342697024345398,
+      "eval_runtime": 235.4774,
+      "eval_samples_per_second": 35.413,
+      "eval_steps_per_second": 4.429,
+      "eval_wer": 0.3095554537121907,
+      "step": 6800
+    },
+    {
+      "epoch": 53.44,
+      "learning_rate": 0.00014086153846153845,
+      "loss": 0.3092,
+      "step": 7000
+    },
+    {
+      "epoch": 53.44,
+      "eval_loss": 0.3224042057991028,
+      "eval_runtime": 237.4291,
+      "eval_samples_per_second": 35.122,
+      "eval_steps_per_second": 4.393,
+      "eval_wer": 0.302910174152154,
+      "step": 7000
+    },
+    {
+      "epoch": 54.96,
+      "learning_rate": 0.00013624615384615383,
+      "loss": 0.2995,
+      "step": 7200
+    },
+    {
+      "epoch": 54.96,
+      "eval_loss": 0.3204595744609833,
+      "eval_runtime": 236.2646,
+      "eval_samples_per_second": 35.295,
+      "eval_steps_per_second": 4.415,
+      "eval_wer": 0.29851054078826766,
+      "step": 7200
+    },
+    {
+      "epoch": 56.49,
+      "learning_rate": 0.0001316307692307692,
+      "loss": 0.304,
+      "step": 7400
+    },
+    {
+      "epoch": 56.49,
+      "eval_loss": 0.35227909684181213,
+      "eval_runtime": 236.3256,
+      "eval_samples_per_second": 35.286,
+      "eval_steps_per_second": 4.413,
+      "eval_wer": 0.30339138405132904,
+      "step": 7400
+    },
+    {
+      "epoch": 58.02,
+      "learning_rate": 0.0001270153846153846,
+      "loss": 0.2952,
+      "step": 7600
+    },
+    {
+      "epoch": 58.02,
+      "eval_loss": 0.3288583755493164,
+      "eval_runtime": 238.857,
+      "eval_samples_per_second": 34.912,
+      "eval_steps_per_second": 4.367,
+      "eval_wer": 0.2934005499541705,
+      "step": 7600
+    },
+    {
+      "epoch": 59.54,
+      "learning_rate": 0.0001224,
+      "loss": 0.2875,
+      "step": 7800
+    },
+    {
+      "epoch": 59.54,
+      "eval_loss": 0.335005521774292,
+      "eval_runtime": 236.8029,
+      "eval_samples_per_second": 35.215,
+      "eval_steps_per_second": 4.405,
+      "eval_wer": 0.3008020164986251,
+      "step": 7800
+    },
+    {
+      "epoch": 61.07,
+      "learning_rate": 0.00011778461538461537,
+      "loss": 0.2868,
+      "step": 8000
+    },
+    {
+      "epoch": 61.07,
+      "eval_loss": 0.3537150025367737,
+      "eval_runtime": 236.5754,
+      "eval_samples_per_second": 35.249,
+      "eval_steps_per_second": 4.409,
+      "eval_wer": 0.3227314390467461,
+      "step": 8000
+    },
+    {
+      "epoch": 62.6,
+      "learning_rate": 0.00011316923076923076,
+      "loss": 0.2875,
+      "step": 8200
+    },
+    {
+      "epoch": 62.6,
+      "eval_loss": 0.3389277756214142,
+      "eval_runtime": 234.9228,
+      "eval_samples_per_second": 35.497,
+      "eval_steps_per_second": 4.44,
+      "eval_wer": 0.29704399633363887,
+      "step": 8200
+    },
+    {
+      "epoch": 64.12,
+      "learning_rate": 0.00010855384615384616,
+      "loss": 0.2778,
+      "step": 8400
+    },
+    {
+      "epoch": 64.12,
+      "eval_loss": 0.33703726530075073,
+      "eval_runtime": 234.6165,
+      "eval_samples_per_second": 35.543,
+      "eval_steps_per_second": 4.446,
+      "eval_wer": 0.29596700274977084,
+      "step": 8400
+    },
+    {
+      "epoch": 65.65,
+      "learning_rate": 0.00010393846153846154,
+      "loss": 0.2706,
+      "step": 8600
+    },
+    {
+      "epoch": 65.65,
+      "eval_loss": 0.32503727078437805,
+      "eval_runtime": 234.7831,
+      "eval_samples_per_second": 35.518,
+      "eval_steps_per_second": 4.442,
+      "eval_wer": 0.28015582034830433,
+      "step": 8600
+    },
+    {
+      "epoch": 67.18,
+      "learning_rate": 9.934615384615383e-05,
+      "loss": 0.2669,
+      "step": 8800
+    },
+    {
+      "epoch": 67.18,
+      "eval_loss": 0.335059255361557,
+      "eval_runtime": 234.788,
+      "eval_samples_per_second": 35.517,
+      "eval_steps_per_second": 4.442,
+      "eval_wer": 0.2902841429880843,
+      "step": 8800
+    },
+    {
+      "epoch": 68.7,
+      "learning_rate": 9.473076923076922e-05,
+      "loss": 0.2615,
+      "step": 9000
+    },
+    {
+      "epoch": 68.7,
+      "eval_loss": 0.3381515145301819,
+      "eval_runtime": 236.5391,
+      "eval_samples_per_second": 35.254,
+      "eval_steps_per_second": 4.409,
+      "eval_wer": 0.29885426214482125,
+      "step": 9000
+    },
+    {
+      "epoch": 70.23,
+      "learning_rate": 9.01153846153846e-05,
+      "loss": 0.2563,
+      "step": 9200
+    },
+    {
+      "epoch": 70.23,
+      "eval_loss": 0.3312215507030487,
+      "eval_runtime": 236.7043,
+      "eval_samples_per_second": 35.23,
+      "eval_steps_per_second": 4.406,
+      "eval_wer": 0.2974793767186068,
+      "step": 9200
+    },
+    {
+      "epoch": 71.76,
+      "learning_rate": 8.549999999999999e-05,
+      "loss": 0.2546,
+      "step": 9400
+    },
+    {
+      "epoch": 71.76,
+      "eval_loss": 0.3212486505508423,
+      "eval_runtime": 238.4241,
+      "eval_samples_per_second": 34.975,
+      "eval_steps_per_second": 4.375,
+      "eval_wer": 0.30034372135655363,
+      "step": 9400
+    },
+    {
+      "epoch": 73.28,
+      "learning_rate": 8.088461538461537e-05,
+      "loss": 0.2482,
+      "step": 9600
+    },
+    {
+      "epoch": 73.28,
+      "eval_loss": 0.3337170481681824,
+      "eval_runtime": 237.0835,
+      "eval_samples_per_second": 35.173,
+      "eval_steps_per_second": 4.399,
+      "eval_wer": 0.30907424381301557,
+      "step": 9600
+    },
+    {
+      "epoch": 74.81,
+      "learning_rate": 7.626923076923075e-05,
+      "loss": 0.2504,
+      "step": 9800
+    },
+    {
+      "epoch": 74.81,
+      "eval_loss": 0.33080631494522095,
+      "eval_runtime": 236.549,
+      "eval_samples_per_second": 35.253,
+      "eval_steps_per_second": 4.409,
+      "eval_wer": 0.3109761686526123,
+      "step": 9800
+    },
+    {
+      "epoch": 76.34,
+      "learning_rate": 7.165384615384615e-05,
+      "loss": 0.2456,
+      "step": 10000
+    },
+    {
+      "epoch": 76.34,
+      "eval_loss": 0.31574100255966187,
+      "eval_runtime": 235.6572,
+      "eval_samples_per_second": 35.386,
+      "eval_steps_per_second": 4.426,
+      "eval_wer": 0.3117781851512374,
+      "step": 10000
+    },
+    {
+      "epoch": 77.86,
+      "learning_rate": 6.703846153846153e-05,
+      "loss": 0.2363,
+      "step": 10200
+    },
+    {
+      "epoch": 77.86,
+      "eval_loss": 0.3251018524169922,
+      "eval_runtime": 236.3835,
+      "eval_samples_per_second": 35.277,
+      "eval_steps_per_second": 4.412,
+      "eval_wer": 0.31439046746104493,
+      "step": 10200
+    },
+    {
+      "epoch": 79.39,
+      "learning_rate": 6.242307692307691e-05,
+      "loss": 0.2319,
+      "step": 10400
+    },
+    {
+      "epoch": 79.39,
+      "eval_loss": 0.32527828216552734,
+      "eval_runtime": 236.1364,
+      "eval_samples_per_second": 35.314,
+      "eval_steps_per_second": 4.417,
+      "eval_wer": 0.30382676443629697,
+      "step": 10400
+    },
+    {
+      "epoch": 80.92,
+      "learning_rate": 5.7807692307692304e-05,
+      "loss": 0.2266,
+      "step": 10600
+    },
+    {
+      "epoch": 80.92,
+      "eval_loss": 0.3374153673648834,
+      "eval_runtime": 236.6995,
+      "eval_samples_per_second": 35.23,
+      "eval_steps_per_second": 4.406,
+      "eval_wer": 0.30382676443629697,
+      "step": 10600
+    },
+    {
+      "epoch": 82.44,
+      "learning_rate": 5.321538461538461e-05,
+      "loss": 0.2279,
+      "step": 10800
+    },
+    {
+      "epoch": 82.44,
+      "eval_loss": 0.32676786184310913,
+      "eval_runtime": 235.2934,
+      "eval_samples_per_second": 35.441,
+      "eval_steps_per_second": 4.433,
+      "eval_wer": 0.29640238313473877,
+      "step": 10800
+    },
+    {
+      "epoch": 83.97,
+      "learning_rate": 4.8599999999999995e-05,
+      "loss": 0.2231,
+      "step": 11000
+    },
+    {
+      "epoch": 83.97,
+      "eval_loss": 0.32775917649269104,
+      "eval_runtime": 237.2828,
+      "eval_samples_per_second": 35.144,
+      "eval_steps_per_second": 4.396,
+      "eval_wer": 0.29502749770852427,
+      "step": 11000
+    },
+    {
+      "epoch": 85.5,
+      "learning_rate": 4.398461538461538e-05,
+      "loss": 0.2185,
+      "step": 11200
+    },
+    {
+      "epoch": 85.5,
+      "eval_loss": 0.3461829721927643,
+      "eval_runtime": 234.8207,
+      "eval_samples_per_second": 35.512,
+      "eval_steps_per_second": 4.442,
+      "eval_wer": 0.29814390467461044,
+      "step": 11200
+    },
+    {
+      "epoch": 87.02,
+      "learning_rate": 3.9369230769230767e-05,
+      "loss": 0.2245,
+      "step": 11400
+    },
+    {
+      "epoch": 87.02,
+      "eval_loss": 0.3311368525028229,
+      "eval_runtime": 234.6868,
+      "eval_samples_per_second": 35.532,
+      "eval_steps_per_second": 4.444,
+      "eval_wer": 0.2894821264894592,
+      "step": 11400
+    },
+    {
+      "epoch": 88.55,
+      "learning_rate": 3.475384615384615e-05,
+      "loss": 0.223,
+      "step": 11600
+    },
+    {
+      "epoch": 88.55,
+      "eval_loss": 0.3325417935848236,
+      "eval_runtime": 234.9788,
+      "eval_samples_per_second": 35.488,
+      "eval_steps_per_second": 4.439,
+      "eval_wer": 0.2876947754353804,
+      "step": 11600
+    },
+    {
+      "epoch": 90.08,
+      "learning_rate": 3.0138461538461538e-05,
+      "loss": 0.2121,
+      "step": 11800
+    },
+    {
+      "epoch": 90.08,
+      "eval_loss": 0.333699494600296,
+      "eval_runtime": 237.3406,
+      "eval_samples_per_second": 35.135,
+      "eval_steps_per_second": 4.395,
+      "eval_wer": 0.282813932172319,
+      "step": 11800
+    },
+    {
+      "epoch": 91.6,
+      "learning_rate": 2.552307692307692e-05,
+      "loss": 0.2126,
+      "step": 12000
+    },
+    {
+      "epoch": 91.6,
+      "eval_loss": 0.3324645757675171,
+      "eval_runtime": 234.7938,
+      "eval_samples_per_second": 35.516,
+      "eval_steps_per_second": 4.442,
+      "eval_wer": 0.28077451879010085,
+      "step": 12000
+    },
+    {
+      "epoch": 93.13,
+      "learning_rate": 2.0907692307692305e-05,
+      "loss": 0.2027,
+      "step": 12200
+    },
+    {
+      "epoch": 93.13,
+      "eval_loss": 0.3276507258415222,
+      "eval_runtime": 235.1021,
+      "eval_samples_per_second": 35.47,
+      "eval_steps_per_second": 4.436,
+      "eval_wer": 0.2819660861594867,
+      "step": 12200
+    },
+    {
+      "epoch": 94.66,
+      "learning_rate": 1.629230769230769e-05,
+      "loss": 0.2058,
+      "step": 12400
+    },
+    {
+      "epoch": 94.66,
+      "eval_loss": 0.33078595995903015,
+      "eval_runtime": 236.8707,
+      "eval_samples_per_second": 35.205,
+      "eval_steps_per_second": 4.403,
+      "eval_wer": 0.2827451879010083,
+      "step": 12400
+    },
+    {
+      "epoch": 96.18,
+      "learning_rate": 1.1676923076923075e-05,
+      "loss": 0.1991,
+      "step": 12600
+    },
+    {
+      "epoch": 96.18,
+      "eval_loss": 0.3278521001338959,
+      "eval_runtime": 235.6753,
+      "eval_samples_per_second": 35.383,
+      "eval_steps_per_second": 4.426,
+      "eval_wer": 0.2820119156736939,
+      "step": 12600
+    },
+    {
+      "epoch": 97.71,
+      "learning_rate": 7.061538461538461e-06,
+      "loss": 0.1991,
+      "step": 12800
+    },
+    {
+      "epoch": 97.71,
+      "eval_loss": 0.3299693167209625,
+      "eval_runtime": 236.8407,
+      "eval_samples_per_second": 35.209,
+      "eval_steps_per_second": 4.404,
+      "eval_wer": 0.28221814848762605,
+      "step": 12800
+    },
+    {
+      "epoch": 99.24,
+      "learning_rate": 2.446153846153846e-06,
+      "loss": 0.1986,
+      "step": 13000
+    },
+    {
+      "epoch": 99.24,
+      "eval_loss": 0.3284846246242523,
+      "eval_runtime": 236.4429,
+      "eval_samples_per_second": 35.269,
+      "eval_steps_per_second": 4.411,
+      "eval_wer": 0.2834555453712191,
+      "step": 13000
+    },
+    {
+      "epoch": 100.0,
+      "step": 13100,
+      "total_flos": 8.36308166572502e+19,
+      "train_loss": 0.29656382378731067,
+      "train_runtime": 73649.3567,
+      "train_samples_per_second": 34.023,
+      "train_steps_per_second": 0.178
+    }
+  ],
+  "max_steps": 13100,
+  "num_train_epochs": 100,
+  "total_flos": 8.36308166572502e+19,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1c646fe11ac72715f880fbe76c8724ae3a051b5d9a16b634b10373429a53a18
+size 2991

vocab.json

@@ -0,0 +1 @@
+{"-": 1, "a": 2, "b": 3, "c": 4, "d": 5, "e": 6, "f": 7, "g": 8, "h": 9, "i": 10, "j": 11, "k": 12, "l": 13, "m": 14, "n": 15, "o": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "u": 22, "v": 23, "w": 24, "x": 25, "y": 26, "z": 27, "â": 28, "ç": 29, "ë": 30, "î": 31, "ö": 32, "ü": 33, "ğ": 34, "ı": 35, "ş": 36, "̇": 37, "|": 0, "[UNK]": 38, "[PAD]": 39}