Init

.gitattributes

@@ -1,6 +1,7 @@
 *.7z filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text
 *.ftz filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text
@@ -16,12 +17,13 @@
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
 *.rar filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text 
 *.tar.* filter=lfs diff=lfs merge=lfs -text
 *.tflite filter=lfs diff=lfs merge=lfs -text
 *.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
 *.xz filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
+*.arpa filter=lfs diff=lfs merge=lfs -text
+*.txt filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,64 @@
 ---
+language:
+- uk
 license: apache-2.0
+tags:
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_7_0
+- generated_from_trainer
+- uk
+xdatasets:
+- mozilla-foundation/common_voice_7_0
 ---
+
+# Ukrainian STT model (with the Big Language Model formed on News Dataset)
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - UK dataset.
+
+Follow our community in Telegram: https://t.me/speech_recognition_uk 
+
+Attribution to the dataset of Language Model:
+
+- Chaplynskyi, D. et al. (2021) lang-uk Ukrainian Ubercorpus [Data set]. https://lang.org.ua/uk/corpora/#anchor4
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 5e-05
+- train_batch_size: 8
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 20
+- total_train_batch_size: 160
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 500
+- num_epochs: 100.0
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Wer    | Cer    |
+|:-------------:|:-----:|:----:|:---------------:|:------:|:------:|
+| 1.2815        | 7.93  | 500  | 0.3536          | 0.4753 | 0.1009 |
+| 1.0869        | 15.86 | 1000 | 0.2317          | 0.3111 | 0.0614 |
+| 0.9984        | 23.8  | 1500 | 0.2022          | 0.2676 | 0.0521 |
+| 0.975         | 31.74 | 2000 | 0.1948          | 0.2469 | 0.0487 |
+| 0.9306        | 39.67 | 2500 | 0.1916          | 0.2377 | 0.0464 |
+| 0.8868        | 47.61 | 3000 | 0.1903          | 0.2257 | 0.0439 |
+| 0.8424        | 55.55 | 3500 | 0.1786          | 0.2206 | 0.0423 |
+| 0.8126        | 63.49 | 4000 | 0.1849          | 0.2160 | 0.0416 |
+| 0.7901        | 71.42 | 4500 | 0.1869          | 0.2138 | 0.0413 |
+| 0.7671        | 79.36 | 5000 | 0.1855          | 0.2075 | 0.0394 |
+| 0.7467        | 87.3  | 5500 | 0.1884          | 0.2049 | 0.0389 |
+| 0.731         | 95.24 | 6000 | 0.1877          | 0.2060 | 0.0387 |
+
+
+### Framework versions
+
+- Transformers 4.16.0.dev0
+- Pytorch 1.10.1+cu102
+- Datasets 1.18.1.dev0
+- Tokenizers 0.11.0

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 51, "</s>": 52}

all_results.json

@@ -0,0 +1,15 @@
+{
+    "epoch": 99.99,
+    "eval_cer": 0.03843434091927693,
+    "eval_loss": 0.18747110664844513,
+    "eval_runtime": 268.1183,
+    "eval_samples": 4332,
+    "eval_samples_per_second": 16.157,
+    "eval_steps_per_second": 2.021,
+    "eval_wer": 0.20326104163368688,
+    "train_loss": 1.049089940994505,
+    "train_runtime": 95054.1856,
+    "train_samples": 10193,
+    "train_samples_per_second": 10.723,
+    "train_steps_per_second": 0.066
+}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "a", "c", "e", "i", "j", "k", "l", "m", "n", "o", "p", "u", "x", "y", "\u0301", "\u0430", "\u0431", "\u0432", "\u0433", "\u0434", "\u0435", "\u0436", "\u0437", "\u0438", "\u0439", "\u043a", "\u043b", "\u043c", "\u043d", "\u043e", "\u043f", "\u0440", "\u0441", "\u0442", "\u0443", "\u0444", "\u0445", "\u0446", "\u0447", "\u0448", "\u0449", "\u044c", "\u044e", "\u044f", "\u0454", "\u0456", "\u0457", "\u0491", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

config.json

@@ -0,0 +1,107 @@
+{
+  "_name_or_path": "facebook/wav2vec2-xls-r-1b",
+  "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": 1024,
+  "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": 1280,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "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": 48,
+  "num_negatives": 100,
+  "output_hidden_size": 1280,
+  "pad_token_id": 50,
+  "proj_codevector_dim": 1024,
+  "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": "4.16.0.dev0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 53,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+from datasets import load_dataset, load_metric, Audio, Dataset
+from transformers import pipeline, AutoFeatureExtractor
+import re
+import argparse
+import unicodedata
+from typing import Dict
+
+
+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 = '[,?.!\-\;\:\"“%‘”�—’…–]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = text.lower()
+    # normalize non-standard (stylized) unicode characters
+    text = unicodedata.normalize('NFKC', text)
+    # remove punctuation
+    text = re.sub(chars_to_ignore_regex, "", text)
+
+    # Let's also make sure we split on all kinds of newlines, spaces, etc...
+    text = " ".join(text.split())
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, 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
+    asr = pipeline("automatic-speech-recognition", model=args.model_id)
+
+    # 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(
+        "--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 None. For long audio files a good value would be 5.0 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to None. For long audio files a good value would be 1.0 seconds."
+    )
+    parser.add_argument(
+        "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
+    )
+    args = parser.parse_args()
+
+    main(args)

eval_results.json

@@ -0,0 +1,10 @@
+{
+    "epoch": 99.99,
+    "eval_cer": 0.03843434091927693,
+    "eval_loss": 0.18747110664844513,
+    "eval_runtime": 268.1183,
+    "eval_samples": 4332,
+    "eval_samples_per_second": 16.157,
+    "eval_steps_per_second": 2.021,
+    "eval_wer": 0.20326104163368688
+}

language_model/5gram.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:216af45a8218db11be16723d33bc458935d310e7d3852297e51af7447689ab3f
+size 5899238255

language_model/attrs.json

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

language_model/unigrams.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b685f9bd31076e9c32363267e3808a6e17ffd36e90488b360a64f8ffb290c2a
+size 64208069

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 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 sha256:6933edff2d809c0035d6ce0fffe2bf975c5d0d3f2932f2c4545f47a769d04ce4
+size 3850584305

run.sh

@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+
+python run_speech_recognition_ctc.py \
+	--dataset_name="mozilla-foundation/common_voice_7_0" \
+	--model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+	--dataset_config_name="uk" \
+	--output_dir="./cv-uk-ft" \
+	--num_train_epochs="100" \
+	--per_device_train_batch_size="8" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="20" \
+	--learning_rate="5e-5" \
+	--warmup_steps="500" \
+	--length_column_name="input_length" \
+	--evaluation_strategy="steps" \
+	--text_column_name="sentence" \
+	--save_steps="500" \
+	--eval_steps="500" \
+	--logging_steps="50" \
+	--layerdrop="0.0" \
+	--activation_dropout="0.1" \
+	--eval_metrics wer cer \
+	--save_total_limit="3" \
+	--feat_proj_dropout="0.0" \
+	--mask_time_prob="0.75" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.25" \
+	--mask_feature_length="64" \
+	--freeze_feature_encoder \
+	--chars_to_ignore « » ы — – ՚ ’ … , ? . ! - \; \: \" “ % ‘ \` ” � \
+	--fp16 \
+	--group_by_length \
+	--use_auth_token \
+	--report_to="tensorboard" \
+	--do_train --do_eval \
+	--gradient_checkpointing

run_speech_recognition_ctc.py

@@ -0,0 +1,737 @@
+#!/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("4.16.0.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 'train'"
+        },
+    )
+    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]}
+
+    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/Jan26_11-22-42_job-df329b21-d243-4736-8f96-d11192aeb370/1643196546.6195295/events.out.tfevents.1643196546.job-df329b21-d243-4736-8f96-d11192aeb370.13810.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dccefa87dfb9f1a70748307d90324380c39d8b93d9af111a89d51ba81de70d24
+size 4778

runs/Jan26_11-22-42_job-df329b21-d243-4736-8f96-d11192aeb370/events.out.tfevents.1643196546.job-df329b21-d243-4736-8f96-d11192aeb370.13810.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1dd1de0119ae5b997385fc2a7feb994debd200ac3055775715f8355a07eb0d2
+size 29217

runs/Jan26_11-22-42_job-df329b21-d243-4736-8f96-d11192aeb370/events.out.tfevents.1643291875.job-df329b21-d243-4736-8f96-d11192aeb370.13810.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cddb9d99a97b8c4323dfd96a1f6fef25c3c7ff9009b73aa856629224e0ce4325
+size 405

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}, {"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": "final", "tokenizer_class": "Wav2Vec2CTCTokenizer", "processor_class": "Wav2Vec2ProcessorWithLM"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 99.99,
+    "train_loss": 1.049089940994505,
+    "train_runtime": 95054.1856,
+    "train_samples": 10193,
+    "train_samples_per_second": 10.723,
+    "train_steps_per_second": 0.066
+}

trainer_state.json

@@ -0,0 +1,901 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 99.98823529411764,
+  "global_step": 6300,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.78,
+      "learning_rate": 4.9000000000000005e-06,
+      "loss": 11.1133,
+      "step": 50
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 9.900000000000002e-06,
+      "loss": 3.3967,
+      "step": 100
+    },
+    {
+      "epoch": 2.38,
+      "learning_rate": 1.49e-05,
+      "loss": 3.2205,
+      "step": 150
+    },
+    {
+      "epoch": 3.17,
+      "learning_rate": 1.9900000000000003e-05,
+      "loss": 2.8143,
+      "step": 200
+    },
+    {
+      "epoch": 3.96,
+      "learning_rate": 2.4900000000000002e-05,
+      "loss": 1.9249,
+      "step": 250
+    },
+    {
+      "epoch": 4.75,
+      "learning_rate": 2.9900000000000002e-05,
+      "loss": 1.6708,
+      "step": 300
+    },
+    {
+      "epoch": 5.55,
+      "learning_rate": 3.49e-05,
+      "loss": 1.5501,
+      "step": 350
+    },
+    {
+      "epoch": 6.35,
+      "learning_rate": 3.99e-05,
+      "loss": 1.4258,
+      "step": 400
+    },
+    {
+      "epoch": 7.14,
+      "learning_rate": 4.49e-05,
+      "loss": 1.332,
+      "step": 450
+    },
+    {
+      "epoch": 7.93,
+      "learning_rate": 4.99e-05,
+      "loss": 1.2815,
+      "step": 500
+    },
+    {
+      "epoch": 7.93,
+      "eval_cer": 0.10093122852447588,
+      "eval_loss": 0.35359087586402893,
+      "eval_runtime": 257.0651,
+      "eval_samples_per_second": 16.852,
+      "eval_steps_per_second": 2.108,
+      "eval_wer": 0.47525724236188066,
+      "step": 500
+    },
+    {
+      "epoch": 8.72,
+      "learning_rate": 4.957758620689655e-05,
+      "loss": 1.2632,
+      "step": 550
+    },
+    {
+      "epoch": 9.52,
+      "learning_rate": 4.9146551724137934e-05,
+      "loss": 1.2239,
+      "step": 600
+    },
+    {
+      "epoch": 10.31,
+      "learning_rate": 4.871551724137931e-05,
+      "loss": 1.2044,
+      "step": 650
+    },
+    {
+      "epoch": 11.11,
+      "learning_rate": 4.828448275862069e-05,
+      "loss": 1.1918,
+      "step": 700
+    },
+    {
+      "epoch": 11.89,
+      "learning_rate": 4.785344827586207e-05,
+      "loss": 1.1641,
+      "step": 750
+    },
+    {
+      "epoch": 12.69,
+      "learning_rate": 4.742241379310345e-05,
+      "loss": 1.1718,
+      "step": 800
+    },
+    {
+      "epoch": 13.49,
+      "learning_rate": 4.699137931034483e-05,
+      "loss": 1.1638,
+      "step": 850
+    },
+    {
+      "epoch": 14.28,
+      "learning_rate": 4.656034482758621e-05,
+      "loss": 1.1317,
+      "step": 900
+    },
+    {
+      "epoch": 15.08,
+      "learning_rate": 4.612931034482759e-05,
+      "loss": 1.1334,
+      "step": 950
+    },
+    {
+      "epoch": 15.86,
+      "learning_rate": 4.569827586206897e-05,
+      "loss": 1.0869,
+      "step": 1000
+    },
+    {
+      "epoch": 15.86,
+      "eval_cer": 0.06135152631841044,
+      "eval_loss": 0.23165984451770782,
+      "eval_runtime": 262.4285,
+      "eval_samples_per_second": 16.507,
+      "eval_steps_per_second": 2.065,
+      "eval_wer": 0.3110653791356657,
+      "step": 1000
+    },
+    {
+      "epoch": 16.66,
+      "learning_rate": 4.526724137931035e-05,
+      "loss": 1.104,
+      "step": 1050
+    },
+    {
+      "epoch": 17.45,
+      "learning_rate": 4.4836206896551726e-05,
+      "loss": 1.109,
+      "step": 1100
+    },
+    {
+      "epoch": 18.25,
+      "learning_rate": 4.440517241379311e-05,
+      "loss": 1.0902,
+      "step": 1150
+    },
+    {
+      "epoch": 19.05,
+      "learning_rate": 4.397413793103449e-05,
+      "loss": 1.0676,
+      "step": 1200
+    },
+    {
+      "epoch": 19.83,
+      "learning_rate": 4.3543103448275865e-05,
+      "loss": 1.0453,
+      "step": 1250
+    },
+    {
+      "epoch": 20.63,
+      "learning_rate": 4.311206896551725e-05,
+      "loss": 1.0489,
+      "step": 1300
+    },
+    {
+      "epoch": 21.42,
+      "learning_rate": 4.268103448275862e-05,
+      "loss": 1.0495,
+      "step": 1350
+    },
+    {
+      "epoch": 22.22,
+      "learning_rate": 4.2250000000000004e-05,
+      "loss": 1.0325,
+      "step": 1400
+    },
+    {
+      "epoch": 23.02,
+      "learning_rate": 4.181896551724138e-05,
+      "loss": 1.0298,
+      "step": 1450
+    },
+    {
+      "epoch": 23.8,
+      "learning_rate": 4.138793103448276e-05,
+      "loss": 0.9984,
+      "step": 1500
+    },
+    {
+      "epoch": 23.8,
+      "eval_cer": 0.052054180568696776,
+      "eval_loss": 0.20215292274951935,
+      "eval_runtime": 259.0562,
+      "eval_samples_per_second": 16.722,
+      "eval_steps_per_second": 2.092,
+      "eval_wer": 0.26762703815102107,
+      "step": 1500
+    },
+    {
+      "epoch": 24.6,
+      "learning_rate": 4.0956896551724136e-05,
+      "loss": 1.0118,
+      "step": 1550
+    },
+    {
+      "epoch": 25.39,
+      "learning_rate": 4.053448275862069e-05,
+      "loss": 1.0165,
+      "step": 1600
+    },
+    {
+      "epoch": 26.19,
+      "learning_rate": 4.0103448275862074e-05,
+      "loss": 1.0075,
+      "step": 1650
+    },
+    {
+      "epoch": 26.97,
+      "learning_rate": 3.967241379310345e-05,
+      "loss": 1.003,
+      "step": 1700
+    },
+    {
+      "epoch": 27.77,
+      "learning_rate": 3.924137931034483e-05,
+      "loss": 0.9905,
+      "step": 1750
+    },
+    {
+      "epoch": 28.56,
+      "learning_rate": 3.8810344827586206e-05,
+      "loss": 1.0019,
+      "step": 1800
+    },
+    {
+      "epoch": 29.36,
+      "learning_rate": 3.837931034482759e-05,
+      "loss": 1.0085,
+      "step": 1850
+    },
+    {
+      "epoch": 30.16,
+      "learning_rate": 3.794827586206896e-05,
+      "loss": 0.9868,
+      "step": 1900
+    },
+    {
+      "epoch": 30.94,
+      "learning_rate": 3.7517241379310345e-05,
+      "loss": 0.9816,
+      "step": 1950
+    },
+    {
+      "epoch": 31.74,
+      "learning_rate": 3.708620689655173e-05,
+      "loss": 0.975,
+      "step": 2000
+    },
+    {
+      "epoch": 31.74,
+      "eval_cer": 0.04868781435187491,
+      "eval_loss": 0.19483695924282074,
+      "eval_runtime": 259.0795,
+      "eval_samples_per_second": 16.721,
+      "eval_steps_per_second": 2.092,
+      "eval_wer": 0.24688934620864333,
+      "step": 2000
+    },
+    {
+      "epoch": 32.53,
+      "learning_rate": 3.66551724137931e-05,
+      "loss": 0.9552,
+      "step": 2050
+    },
+    {
+      "epoch": 33.33,
+      "learning_rate": 3.6224137931034484e-05,
+      "loss": 0.9649,
+      "step": 2100
+    },
+    {
+      "epoch": 34.13,
+      "learning_rate": 3.5793103448275866e-05,
+      "loss": 0.9632,
+      "step": 2150
+    },
+    {
+      "epoch": 34.91,
+      "learning_rate": 3.536206896551724e-05,
+      "loss": 0.9542,
+      "step": 2200
+    },
+    {
+      "epoch": 35.71,
+      "learning_rate": 3.493103448275862e-05,
+      "loss": 0.9686,
+      "step": 2250
+    },
+    {
+      "epoch": 36.5,
+      "learning_rate": 3.45e-05,
+      "loss": 0.9418,
+      "step": 2300
+    },
+    {
+      "epoch": 37.3,
+      "learning_rate": 3.406896551724138e-05,
+      "loss": 0.9295,
+      "step": 2350
+    },
+    {
+      "epoch": 38.09,
+      "learning_rate": 3.363793103448276e-05,
+      "loss": 0.9372,
+      "step": 2400
+    },
+    {
+      "epoch": 38.88,
+      "learning_rate": 3.320689655172414e-05,
+      "loss": 0.9205,
+      "step": 2450
+    },
+    {
+      "epoch": 39.67,
+      "learning_rate": 3.277586206896552e-05,
+      "loss": 0.9306,
+      "step": 2500
+    },
+    {
+      "epoch": 39.67,
+      "eval_cer": 0.046377172451571136,
+      "eval_loss": 0.19161736965179443,
+      "eval_runtime": 258.3157,
+      "eval_samples_per_second": 16.77,
+      "eval_steps_per_second": 2.098,
+      "eval_wer": 0.2377394332752889,
+      "step": 2500
+    },
+    {
+      "epoch": 40.47,
+      "learning_rate": 3.23448275862069e-05,
+      "loss": 0.9331,
+      "step": 2550
+    },
+    {
+      "epoch": 41.27,
+      "learning_rate": 3.1913793103448276e-05,
+      "loss": 0.8936,
+      "step": 2600
+    },
+    {
+      "epoch": 42.06,
+      "learning_rate": 3.148275862068966e-05,
+      "loss": 0.8987,
+      "step": 2650
+    },
+    {
+      "epoch": 42.85,
+      "learning_rate": 3.105172413793104e-05,
+      "loss": 0.8853,
+      "step": 2700
+    },
+    {
+      "epoch": 43.64,
+      "learning_rate": 3.0620689655172415e-05,
+      "loss": 0.9106,
+      "step": 2750
+    },
+    {
+      "epoch": 44.44,
+      "learning_rate": 3.0189655172413794e-05,
+      "loss": 0.8932,
+      "step": 2800
+    },
+    {
+      "epoch": 45.24,
+      "learning_rate": 2.9758620689655176e-05,
+      "loss": 0.9096,
+      "step": 2850
+    },
+    {
+      "epoch": 46.03,
+      "learning_rate": 2.932758620689655e-05,
+      "loss": 0.8919,
+      "step": 2900
+    },
+    {
+      "epoch": 46.82,
+      "learning_rate": 2.8896551724137933e-05,
+      "loss": 0.8744,
+      "step": 2950
+    },
+    {
+      "epoch": 47.61,
+      "learning_rate": 2.8465517241379315e-05,
+      "loss": 0.8868,
+      "step": 3000
+    },
+    {
+      "epoch": 47.61,
+      "eval_cer": 0.04391713560081669,
+      "eval_loss": 0.19031885266304016,
+      "eval_runtime": 265.4438,
+      "eval_samples_per_second": 16.32,
+      "eval_steps_per_second": 2.042,
+      "eval_wer": 0.2257400664872566,
+      "step": 3000
+    },
+    {
+      "epoch": 48.41,
+      "learning_rate": 2.803448275862069e-05,
+      "loss": 0.8793,
+      "step": 3050
+    },
+    {
+      "epoch": 49.2,
+      "learning_rate": 2.7603448275862072e-05,
+      "loss": 0.8739,
+      "step": 3100
+    },
+    {
+      "epoch": 49.99,
+      "learning_rate": 2.717241379310345e-05,
+      "loss": 0.8696,
+      "step": 3150
+    },
+    {
+      "epoch": 50.78,
+      "learning_rate": 2.674137931034483e-05,
+      "loss": 0.863,
+      "step": 3200
+    },
+    {
+      "epoch": 51.58,
+      "learning_rate": 2.6310344827586207e-05,
+      "loss": 0.8612,
+      "step": 3250
+    },
+    {
+      "epoch": 52.38,
+      "learning_rate": 2.587931034482759e-05,
+      "loss": 0.8639,
+      "step": 3300
+    },
+    {
+      "epoch": 53.17,
+      "learning_rate": 2.5448275862068964e-05,
+      "loss": 0.8523,
+      "step": 3350
+    },
+    {
+      "epoch": 53.96,
+      "learning_rate": 2.5017241379310346e-05,
+      "loss": 0.8577,
+      "step": 3400
+    },
+    {
+      "epoch": 54.75,
+      "learning_rate": 2.4586206896551725e-05,
+      "loss": 0.8465,
+      "step": 3450
+    },
+    {
+      "epoch": 55.55,
+      "learning_rate": 2.4155172413793103e-05,
+      "loss": 0.8424,
+      "step": 3500
+    },
+    {
+      "epoch": 55.55,
+      "eval_cer": 0.042293710472586024,
+      "eval_loss": 0.17861121892929077,
+      "eval_runtime": 259.0288,
+      "eval_samples_per_second": 16.724,
+      "eval_steps_per_second": 2.092,
+      "eval_wer": 0.22061104954883648,
+      "step": 3500
+    },
+    {
+      "epoch": 56.35,
+      "learning_rate": 2.3724137931034485e-05,
+      "loss": 0.8436,
+      "step": 3550
+    },
+    {
+      "epoch": 57.14,
+      "learning_rate": 2.3293103448275864e-05,
+      "loss": 0.8404,
+      "step": 3600
+    },
+    {
+      "epoch": 57.93,
+      "learning_rate": 2.2862068965517242e-05,
+      "loss": 0.8304,
+      "step": 3650
+    },
+    {
+      "epoch": 58.72,
+      "learning_rate": 2.2431034482758624e-05,
+      "loss": 0.8331,
+      "step": 3700
+    },
+    {
+      "epoch": 59.52,
+      "learning_rate": 2.2000000000000003e-05,
+      "loss": 0.824,
+      "step": 3750
+    },
+    {
+      "epoch": 60.31,
+      "learning_rate": 2.1568965517241378e-05,
+      "loss": 0.8328,
+      "step": 3800
+    },
+    {
+      "epoch": 61.11,
+      "learning_rate": 2.113793103448276e-05,
+      "loss": 0.8234,
+      "step": 3850
+    },
+    {
+      "epoch": 61.89,
+      "learning_rate": 2.070689655172414e-05,
+      "loss": 0.8098,
+      "step": 3900
+    },
+    {
+      "epoch": 62.69,
+      "learning_rate": 2.0275862068965517e-05,
+      "loss": 0.8287,
+      "step": 3950
+    },
+    {
+      "epoch": 63.49,
+      "learning_rate": 1.98448275862069e-05,
+      "loss": 0.8126,
+      "step": 4000
+    },
+    {
+      "epoch": 63.49,
+      "eval_cer": 0.04164135252228475,
+      "eval_loss": 0.18486249446868896,
+      "eval_runtime": 261.7127,
+      "eval_samples_per_second": 16.553,
+      "eval_steps_per_second": 2.071,
+      "eval_wer": 0.2159886021845813,
+      "step": 4000
+    },
+    {
+      "epoch": 64.28,
+      "learning_rate": 1.9413793103448277e-05,
+      "loss": 0.8089,
+      "step": 4050
+    },
+    {
+      "epoch": 65.08,
+      "learning_rate": 1.8982758620689656e-05,
+      "loss": 0.8126,
+      "step": 4100
+    },
+    {
+      "epoch": 65.86,
+      "learning_rate": 1.8551724137931034e-05,
+      "loss": 0.7975,
+      "step": 4150
+    },
+    {
+      "epoch": 66.66,
+      "learning_rate": 1.8120689655172416e-05,
+      "loss": 0.8049,
+      "step": 4200
+    },
+    {
+      "epoch": 67.45,
+      "learning_rate": 1.7698275862068966e-05,
+      "loss": 0.8088,
+      "step": 4250
+    },
+    {
+      "epoch": 68.25,
+      "learning_rate": 1.7267241379310344e-05,
+      "loss": 0.8038,
+      "step": 4300
+    },
+    {
+      "epoch": 69.05,
+      "learning_rate": 1.6836206896551726e-05,
+      "loss": 0.7886,
+      "step": 4350
+    },
+    {
+      "epoch": 69.83,
+      "learning_rate": 1.6405172413793105e-05,
+      "loss": 0.7735,
+      "step": 4400
+    },
+    {
+      "epoch": 70.63,
+      "learning_rate": 1.5974137931034483e-05,
+      "loss": 0.7837,
+      "step": 4450
+    },
+    {
+      "epoch": 71.42,
+      "learning_rate": 1.5543103448275865e-05,
+      "loss": 0.7901,
+      "step": 4500
+    },
+    {
+      "epoch": 71.42,
+      "eval_cer": 0.04126786514615806,
+      "eval_loss": 0.18691900372505188,
+      "eval_runtime": 261.5464,
+      "eval_samples_per_second": 16.563,
+      "eval_steps_per_second": 2.072,
+      "eval_wer": 0.21383568149438023,
+      "step": 4500
+    },
+    {
+      "epoch": 72.22,
+      "learning_rate": 1.5112068965517242e-05,
+      "loss": 0.7949,
+      "step": 4550
+    },
+    {
+      "epoch": 73.02,
+      "learning_rate": 1.468103448275862e-05,
+      "loss": 0.7893,
+      "step": 4600
+    },
+    {
+      "epoch": 73.8,
+      "learning_rate": 1.4249999999999999e-05,
+      "loss": 0.7603,
+      "step": 4650
+    },
+    {
+      "epoch": 74.6,
+      "learning_rate": 1.3818965517241381e-05,
+      "loss": 0.776,
+      "step": 4700
+    },
+    {
+      "epoch": 75.39,
+      "learning_rate": 1.338793103448276e-05,
+      "loss": 0.7755,
+      "step": 4750
+    },
+    {
+      "epoch": 76.19,
+      "learning_rate": 1.2956896551724138e-05,
+      "loss": 0.7751,
+      "step": 4800
+    },
+    {
+      "epoch": 76.97,
+      "learning_rate": 1.2525862068965518e-05,
+      "loss": 0.7608,
+      "step": 4850
+    },
+    {
+      "epoch": 77.77,
+      "learning_rate": 1.2094827586206897e-05,
+      "loss": 0.7663,
+      "step": 4900
+    },
+    {
+      "epoch": 78.56,
+      "learning_rate": 1.1663793103448277e-05,
+      "loss": 0.7656,
+      "step": 4950
+    },
+    {
+      "epoch": 79.36,
+      "learning_rate": 1.1232758620689656e-05,
+      "loss": 0.7671,
+      "step": 5000
+    },
+    {
+      "epoch": 79.36,
+      "eval_cer": 0.03937054927543449,
+      "eval_loss": 0.18550464510917664,
+      "eval_runtime": 260.3539,
+      "eval_samples_per_second": 16.639,
+      "eval_steps_per_second": 2.082,
+      "eval_wer": 0.20747190121893302,
+      "step": 5000
+    },
+    {
+      "epoch": 80.16,
+      "learning_rate": 1.0801724137931036e-05,
+      "loss": 0.7694,
+      "step": 5050
+    },
+    {
+      "epoch": 80.94,
+      "learning_rate": 1.0370689655172414e-05,
+      "loss": 0.7672,
+      "step": 5100
+    },
+    {
+      "epoch": 81.74,
+      "learning_rate": 9.939655172413793e-06,
+      "loss": 0.7444,
+      "step": 5150
+    },
+    {
+      "epoch": 82.53,
+      "learning_rate": 9.508620689655173e-06,
+      "loss": 0.7534,
+      "step": 5200
+    },
+    {
+      "epoch": 83.33,
+      "learning_rate": 9.077586206896552e-06,
+      "loss": 0.7453,
+      "step": 5250
+    },
+    {
+      "epoch": 84.13,
+      "learning_rate": 8.646551724137932e-06,
+      "loss": 0.7494,
+      "step": 5300
+    },
+    {
+      "epoch": 84.91,
+      "learning_rate": 8.224137931034483e-06,
+      "loss": 0.7425,
+      "step": 5350
+    },
+    {
+      "epoch": 85.71,
+      "learning_rate": 7.793103448275863e-06,
+      "loss": 0.7499,
+      "step": 5400
+    },
+    {
+      "epoch": 86.5,
+      "learning_rate": 7.370689655172413e-06,
+      "loss": 0.735,
+      "step": 5450
+    },
+    {
+      "epoch": 87.3,
+      "learning_rate": 6.939655172413794e-06,
+      "loss": 0.7467,
+      "step": 5500
+    },
+    {
+      "epoch": 87.3,
+      "eval_cer": 0.03894228375080922,
+      "eval_loss": 0.18841499090194702,
+      "eval_runtime": 261.43,
+      "eval_samples_per_second": 16.57,
+      "eval_steps_per_second": 2.073,
+      "eval_wer": 0.20490739274972297,
+      "step": 5500
+    },
+    {
+      "epoch": 88.09,
+      "learning_rate": 6.508620689655173e-06,
+      "loss": 0.7348,
+      "step": 5550
+    },
+    {
+      "epoch": 88.88,
+      "learning_rate": 6.0775862068965515e-06,
+      "loss": 0.7244,
+      "step": 5600
+    },
+    {
+      "epoch": 89.67,
+      "learning_rate": 5.646551724137932e-06,
+      "loss": 0.7394,
+      "step": 5650
+    },
+    {
+      "epoch": 90.47,
+      "learning_rate": 5.21551724137931e-06,
+      "loss": 0.7423,
+      "step": 5700
+    },
+    {
+      "epoch": 91.27,
+      "learning_rate": 4.78448275862069e-06,
+      "loss": 0.7251,
+      "step": 5750
+    },
+    {
+      "epoch": 92.06,
+      "learning_rate": 4.353448275862069e-06,
+      "loss": 0.7304,
+      "step": 5800
+    },
+    {
+      "epoch": 92.85,
+      "learning_rate": 3.9224137931034484e-06,
+      "loss": 0.7153,
+      "step": 5850
+    },
+    {
+      "epoch": 93.64,
+      "learning_rate": 3.491379310344828e-06,
+      "loss": 0.7287,
+      "step": 5900
+    },
+    {
+      "epoch": 94.44,
+      "learning_rate": 3.0603448275862068e-06,
+      "loss": 0.7349,
+      "step": 5950
+    },
+    {
+      "epoch": 95.24,
+      "learning_rate": 2.6293103448275866e-06,
+      "loss": 0.731,
+      "step": 6000
+    },
+    {
+      "epoch": 95.24,
+      "eval_cer": 0.03871819132513321,
+      "eval_loss": 0.1877404898405075,
+      "eval_runtime": 259.3367,
+      "eval_samples_per_second": 16.704,
+      "eval_steps_per_second": 2.09,
+      "eval_wer": 0.2059838530948235,
+      "step": 6000
+    },
+    {
+      "epoch": 96.03,
+      "learning_rate": 2.1982758620689655e-06,
+      "loss": 0.7151,
+      "step": 6050
+    },
+    {
+      "epoch": 96.82,
+      "learning_rate": 1.7672413793103449e-06,
+      "loss": 0.713,
+      "step": 6100
+    },
+    {
+      "epoch": 97.61,
+      "learning_rate": 1.3362068965517243e-06,
+      "loss": 0.7257,
+      "step": 6150
+    },
+    {
+      "epoch": 98.41,
+      "learning_rate": 9.051724137931035e-07,
+      "loss": 0.7287,
+      "step": 6200
+    },
+    {
+      "epoch": 99.2,
+      "learning_rate": 4.7413793103448276e-07,
+      "loss": 0.7273,
+      "step": 6250
+    },
+    {
+      "epoch": 99.99,
+      "learning_rate": 4.310344827586207e-08,
+      "loss": 0.7082,
+      "step": 6300
+    },
+    {
+      "epoch": 99.99,
+      "step": 6300,
+      "total_flos": 4.0887685530877926e+20,
+      "train_loss": 1.049089940994505,
+      "train_runtime": 95054.1856,
+      "train_samples_per_second": 10.723,
+      "train_steps_per_second": 0.066
+    }
+  ],
+  "max_steps": 6300,
+  "num_train_epochs": 100,
+  "total_flos": 4.0887685530877926e+20,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d28ad45d3721baa2d9d1ae37754a308d7348290cd7784d5d9c00aa9c1ffe0061
+size 3055

vocab.json

@@ -0,0 +1 @@
+{"a": 1, "c": 2, "e": 3, "i": 4, "j": 5, "k": 6, "l": 7, "m": 8, "n": 9, "o": 10, "p": 11, "u": 12, "x": 13, "y": 14, "́": 15, "а": 16, "б": 17, "в": 18, "г": 19, "д": 20, "е": 21, "ж": 22, "з": 23, "и": 24, "й": 25, "к": 26, "л": 27, "м": 28, "н": 29, "о": 30, "п": 31, "р": 32, "с": 33, "т": 34, "у": 35, "ф": 36, "х": 37, "ц": 38, "ч": 39, "ш": 40, "щ": 41, "ь": 42, "ю": 43, "я": 44, "є": 45, "і": 46, "ї": 47, "ґ": 48, "|": 0, "[UNK]": 49, "[PAD]": 50}