Add checkpoint

README.md

@@ -0,0 +1,69 @@
+---
+language: ru
+datasets:
+- Common Voice 
+metrics:
+- wer
+tags:
+- audio
+- speech
+- wav2vec2
+- ru
+- russian-speech-corpus
+- automatic-speech-recognition
+- speech
+- PyTorch
+license: apache-2.0
+model-index:
+- name: Edresson Casanova Wav2vec2 Large 100k Voxpopuli fine-tuned with  Common Voice and M-AILABS in Russian 
+  results:
+  - task: 
+      name: Speech Recognition
+      type: automatic-speech-recognition
+    metrics:
+       - name: Test Common Voice 7.0 WER
+         type: wer
+         value: 24.80
+---
+
+# Wav2vec2 Large 100k Voxpopuli fine-tuned with Common Voice and M-AILABS in Russian 
+
+[Wav2vec2 Large 100k Voxpopuli](https://huggingface.co/facebook/wav2vec2-large-100k-voxpopuli) fine-tuned in Russian using the Common Voice 7.0 and M-AILABS.
+
+
+
+# Use this model
+
+```python
+
+from transformers import AutoTokenizer, Wav2Vec2ForCTC
+  
+tokenizer = AutoTokenizer.from_pretrained("Edresson/wav2vec2-large-100k-voxpopuli-ft-Common-Voice_plus_TTS-Dataset-russian")
+
+model = Wav2Vec2ForCTC.from_pretrained("Edresson/wav2vec2-large-100k-voxpopuli-ft-Common-Voice_plus_TTS-Dataset-russian")
+```
+# Results
+For the results check the [article (Soon)]()
+
+# Example test with Common Voice Dataset
+
+
+```python
+dataset = load_dataset("common_voice", "pt", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
+
+resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
+
+def map_to_array(batch):
+    speech, _ = torchaudio.load(batch["path"])
+    batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
+    batch["sampling_rate"] = resampler.new_freq
+    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
+    return batch
+```
+
+```python
+ds = dataset.map(map_to_array)
+result = ds.map(map_to_pred, batched=True, batch_size=1, remove_columns=list(ds.features.keys()))
+print(wer.compute(predictions=result["predicted"], references=result["target"]))
+```
+

all_results.json

@@ -0,0 +1,11 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.44034817814826965,
+    "eval_runtime": 322.1126,
+    "eval_samples": 8422,
+    "eval_samples_per_second": 26.146,
+    "eval_wer": 0.36474384728036185,
+    "train_runtime": 215429.416,
+    "train_samples": 23875,
+    "train_samples_per_second": 0.058
+}

config.json

@@ -0,0 +1,84 @@
+{
+  "_name_or_path": "facebook/wav2vec2-large-100k-voxpopuli",
+  "activation_dropout": 0.0,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "codevector_dim": 768,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": true,
+  "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": 0.1,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "gradient_checkpointing": true,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.0,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "mask_feature_length": 10,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.05,
+  "mask_time_selection": "static",
+  "model_type": "wav2vec2",
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "num_negatives": 100,
+  "pad_token_id": 0,
+  "proj_codevector_dim": 768,
+  "transformers_version": "4.6.1",
+  "vocab_size": 39
+}

config_train.json

@@ -0,0 +1,171 @@
+{
+    "run_name": "Wav2Vec-fine-tuning-TEDx",
+    "run_description": "Fine tuning TEDx",
+    "seed": 42,
+    // AUDIO PARAMS
+    "sampling_rate": 16000,
+
+    // VOCABULARY PARAMETERS
+    "vocab":{
+        "vocab_path": "example/vocab_example_ru.json", // generic vocab for Portuguese
+        "blank": "<pad>", // blank token for padding
+        "silence": "|", // token between words
+        "unk": "<unk>" // unk token
+    },
+
+    // TRAINING
+    "batch_size": 8,       // Batch size for training.
+    "mixed_precision": true,     // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
+    "early_stop_epochs": 10, // If 0 disabled else Number of epochs for stop training with validation loss dont decrease 
+    "preprocess_dataset": false, // if true, the dataset will be pre-processed and saved in disk, otherwise the audio files will be loaded in each step. Preprocessing makes training faster, but requires much more disk space.
+
+    // OPTIMIZER
+    "epochs": 100,                // total number of epochs to train.
+    "lr": 0.00003,                  // Initial learning rate.
+    "gradient_accumulation_steps": 24, 
+
+    // LOGGING
+    "logging_steps": 100,    // Number of steps to plot.
+    "load_best_model_at_end": true,
+    "save_total_limit": 3,
+    "warmup_ratio": 0.04761904762142857, // 0 disable Ratio of total training steps used for a linear warmup from 0 to learning_rate
+    "warmup_steps": 0, // 0 disable  Number of steps used for a linear warmup from 0 to learning_rate
+
+    // DATA LOADING
+    "num_loader_workers": 8,        // number of training data loader processes. Don't set it too big. 4-8 are goo
+    
+    // MODEL
+    "freeze_feature_extractor": true, // Whether to freeze the feature extractor layers of the model.
+    "attention_dropout": 0.1, // The dropout ratio for the attention probabilities.
+    "activation_dropout": 0.1, // The dropout ratio for activations inside the fully connected layer.
+    "hidden_dropout": 0.1, // The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+    "feat_proj_dropout": 0.1, // The dropout probabilitiy for all 1D convolutional layers in feature extractor.
+    "mask_time_prob": 0.05, //  Propability of each feature vector along the time axis to be chosen as the start of the vector span to be masked.
+    "layerdrop": 0.0,  // The LayerDrop probability. 
+    "gradient_checkpointing": true,  // If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+
+    // ToDo: Implement Time mask and Frequency Mask
+    "audio_augmentation":[
+        // additive noise and room impulse response (RIR) simulation similar to: https://arxiv.org/pdf/2009.14153.pdf
+        {
+            "name": "additive",
+            "sounds_path":"../../datasets/musan/speech/", // download: https://www.openslr.org/17/
+            "lru_cache_size": 32, // Maximum size of the LRU cache for storing noise files in memory
+            "min_snr_in_db": 13.0,
+            "max_snr_in_db": 20.0,
+            // "sample_rate": 16000,
+            "p": 0.25
+        },
+        {
+            "name": "additive",
+            "sounds_path":"../../datasets/musan/music/", // download: https://www.openslr.org/17/
+            "lru_cache_size": 32, // Maximum size of the LRU cache for storing noise files in memory
+            "min_snr_in_db": 5.0,
+            "max_snr_in_db": 15.0,
+            // "sample_rate": 16000,
+            "p": 0.25
+        },
+        {
+            "name": "additive",
+            "sounds_path":"../../datasets/musan/noise/", // download: https://www.openslr.org/17/
+            "lru_cache_size": 32, // Maximum size of the LRU cache for storing noise files in memory
+            "min_snr_in_db": 0.0,
+            "max_snr_in_db": 15.0,
+            // "sample_rate": 16000,
+            "p": 0.25
+        },
+        // rir filter proposed by: https://ieeexplore.ieee.org/document/7953152
+        {
+            "name": "rir",
+            "ir_path": "../../datasets/RIRS_NOISES/simulated_rirs/", // download: https://www.openslr.org/28/
+            "lru_cache_size": 128, // Maximum size of the LRU cache for storing noise files in memory
+            // "sample_rate": 16000,
+            "p": 0.25
+        }
+        , 
+        // {
+        //     "name": "gain",
+        //     "min_gain_in_db": -18.0,
+        //     "max_gain_in_db": 6,
+        //     "p": 0.25 // propability of apply this method, 0 is disable
+        // },
+        {
+            "name": "pitch_shift",
+            "min_semitones": -4,
+            "max_semitones": 4,
+            "p": 0.25 // propability of apply this method, 0 is disable
+        },
+        {
+            "name": "gaussian",
+            "min_amplitude": 0.0001,
+            "max_amplitude": 0.001,
+            "p": 0.25 // propability of apply this method, 0 is disable
+        }
+    ],
+
+    // PATHS
+    "output_path": "../checkpoints/Wav2Vec-voxpopuli/one-speaker/Final-paper/GT/RU/100-epoch/",
+    // CACHE
+    "dataset_cache": "../datasets/",
+
+    // DATASETS
+    "datasets":{ 
+        
+        "files_path": "/workspace/edresson/datasets/Common_Voice/cv-corpus-7.0-2021-07-21/ru/", // relative path for audios It's will be join with the CS
+        "train":
+            [
+               // this dicts is pass directly for the load dataset see the documentation: https://huggingface.co/docs/datasets/package_reference/loading_methods.html#datasets.load_dataset 
+            {
+                "name": "csv",
+                "path": "csv",
+                "data_files": ["/workspace/edresson/datasets/Common_Voice/cv-corpus-7.0-2021-07-21/ru/train_converted.csv"], // csv files
+                "text_column": "text",
+                "path_column": "file_path"
+            }, 
+            {                
+                "name": "csv",
+                "path": "csv",
+                "data_files": ["/workspace/edresson/datasets/M-AILABS/ru_RU/train_converted.csv"], // csv files
+                "text_column": "text",
+                "path_column": "file_path"
+            }
+            ]
+        ,
+        "devel":
+            [ 
+                {
+                    "name": "csv",
+                    "path": "csv",
+                    "data_files": ["/workspace/edresson/datasets/Common_Voice/cv-corpus-7.0-2021-07-21/ru/dev_converted.csv"], // csv files
+                    "text_column": "text",
+                    "path_column": "file_path"
+                }
+        
+            ] 
+            ,
+        "test":
+            {
+                "name": "csv",
+                "path": "csv",
+                "data_files": ["/workspace/edresson/datasets/Common_Voice/cv-corpus-7.0-2021-07-21/ru/test_converted.csv"], // csv files
+                "text_column": "text",
+                "path_column": "file_path"
+            }
+    
+    }//,
+    // used only for test 
+    // "KenLM":{
+    // "kenlm_model_path": "../../kenLM/binaries/subtitle/4-gram/lm.binary", // Path for KenLM model
+    // "lexicon_path": "example/lexicon.lst", // file with all words for limit the decoder search
+    // "beam": 2048, 
+    // "nbest": 1,
+    // "beam_threshold": 25, 
+    // "lm_weight": 1, 
+    // "word_score": -1,
+    // "sil_weight": 0
+    // }
+
+
+
+}
+

eval_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 100.0,
+    "eval_loss": 0.44034817814826965,
+    "eval_runtime": 322.1126,
+    "eval_samples": 8422,
+    "eval_samples_per_second": 26.146,
+    "eval_wer": 0.36474384728036185
+}

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d681c41aa0f502474355e397a9259f32ce1cc6ce698d831c3d75cc170b1912e1
+size 1262089108

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}

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": "|"}

train_results.json

@@ -0,0 +1,6 @@
+{
+    "epoch": 100.0,
+    "train_runtime": 215429.416,
+    "train_samples": 23875,
+    "train_samples_per_second": 0.058
+}

trainer_state.json

@@ -0,0 +1,1573 @@
+{
+  "best_metric": 0.43025869131088257,
+  "best_model_checkpoint": "../checkpoints/Wav2Vec-voxpopuli/one-speaker/Final-paper/GT/RU/100-epoch/checkpoint-12152",
+  "epoch": 99.99698492462312,
+  "global_step": 12400,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "learning_rate": 5.076142131979695e-08,
+      "loss": 11.6061,
+      "step": 1
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 4.974619289340102e-06,
+      "loss": 14.721,
+      "step": 100
+    },
+    {
+      "epoch": 1.0,
+      "eval_loss": 10.616052627563477,
+      "eval_runtime": 323.0492,
+      "eval_samples_per_second": 26.07,
+      "eval_wer": 0.9999740801700341,
+      "step": 124
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 10.5538,
+      "step": 200
+    },
+    {
+      "epoch": 2.0,
+      "eval_loss": 5.4549055099487305,
+      "eval_runtime": 322.5611,
+      "eval_samples_per_second": 26.11,
+      "eval_wer": 1.0,
+      "step": 248
+    },
+    {
+      "epoch": 2.42,
+      "learning_rate": 1.5076142131979694e-05,
+      "loss": 6.6454,
+      "step": 300
+    },
+    {
+      "epoch": 3.0,
+      "eval_loss": 4.0469207763671875,
+      "eval_runtime": 322.6151,
+      "eval_samples_per_second": 26.105,
+      "eval_wer": 1.0,
+      "step": 372
+    },
+    {
+      "epoch": 3.23,
+      "learning_rate": 2.015228426395939e-05,
+      "loss": 4.8946,
+      "step": 400
+    },
+    {
+      "epoch": 4.0,
+      "eval_loss": 3.3774263858795166,
+      "eval_runtime": 321.7327,
+      "eval_samples_per_second": 26.177,
+      "eval_wer": 1.0,
+      "step": 496
+    },
+    {
+      "epoch": 4.03,
+      "learning_rate": 2.522842639593909e-05,
+      "loss": 3.8757,
+      "step": 500
+    },
+    {
+      "epoch": 4.84,
+      "learning_rate": 2.9984757388432552e-05,
+      "loss": 3.3388,
+      "step": 600
+    },
+    {
+      "epoch": 5.0,
+      "eval_loss": 3.1716513633728027,
+      "eval_runtime": 320.8269,
+      "eval_samples_per_second": 26.251,
+      "eval_wer": 1.0,
+      "step": 620
+    },
+    {
+      "epoch": 5.64,
+      "learning_rate": 2.973071386230841e-05,
+      "loss": 3.1981,
+      "step": 700
+    },
+    {
+      "epoch": 6.0,
+      "eval_loss": 3.110382080078125,
+      "eval_runtime": 321.1936,
+      "eval_samples_per_second": 26.221,
+      "eval_wer": 1.0,
+      "step": 744
+    },
+    {
+      "epoch": 6.45,
+      "learning_rate": 2.9476670336184266e-05,
+      "loss": 3.1373,
+      "step": 800
+    },
+    {
+      "epoch": 7.0,
+      "eval_loss": 3.060086727142334,
+      "eval_runtime": 321.2282,
+      "eval_samples_per_second": 26.218,
+      "eval_wer": 1.0,
+      "step": 868
+    },
+    {
+      "epoch": 7.26,
+      "learning_rate": 2.9222626810060127e-05,
+      "loss": 3.0877,
+      "step": 900
+    },
+    {
+      "epoch": 8.0,
+      "eval_loss": 3.034672498703003,
+      "eval_runtime": 321.5996,
+      "eval_samples_per_second": 26.188,
+      "eval_wer": 1.0,
+      "step": 992
+    },
+    {
+      "epoch": 8.06,
+      "learning_rate": 2.8968583283935984e-05,
+      "loss": 3.0573,
+      "step": 1000
+    },
+    {
+      "epoch": 8.87,
+      "learning_rate": 2.8714539757811837e-05,
+      "loss": 3.0092,
+      "step": 1100
+    },
+    {
+      "epoch": 9.0,
+      "eval_loss": 2.9337708950042725,
+      "eval_runtime": 321.5622,
+      "eval_samples_per_second": 26.191,
+      "eval_wer": 1.0,
+      "step": 1116
+    },
+    {
+      "epoch": 9.68,
+      "learning_rate": 2.8460496231687698e-05,
+      "loss": 2.8881,
+      "step": 1200
+    },
+    {
+      "epoch": 10.0,
+      "eval_loss": 2.485553741455078,
+      "eval_runtime": 322.8144,
+      "eval_samples_per_second": 26.089,
+      "eval_wer": 0.9980430528375733,
+      "step": 1240
+    },
+    {
+      "epoch": 10.48,
+      "learning_rate": 2.8206452705563555e-05,
+      "loss": 2.6025,
+      "step": 1300
+    },
+    {
+      "epoch": 11.0,
+      "eval_loss": 1.928577184677124,
+      "eval_runtime": 322.3486,
+      "eval_samples_per_second": 26.127,
+      "eval_wer": 0.9831909902671039,
+      "step": 1364
+    },
+    {
+      "epoch": 11.29,
+      "learning_rate": 2.795240917943941e-05,
+      "loss": 2.2005,
+      "step": 1400
+    },
+    {
+      "epoch": 12.0,
+      "eval_loss": 1.6449922323226929,
+      "eval_runtime": 322.269,
+      "eval_samples_per_second": 26.133,
+      "eval_wer": 0.9675483728826739,
+      "step": 1488
+    },
+    {
+      "epoch": 12.1,
+      "learning_rate": 2.769836565331527e-05,
+      "loss": 1.9289,
+      "step": 1500
+    },
+    {
+      "epoch": 12.9,
+      "learning_rate": 2.7444322127191126e-05,
+      "loss": 1.7455,
+      "step": 1600
+    },
+    {
+      "epoch": 13.0,
+      "eval_loss": 1.4952174425125122,
+      "eval_runtime": 324.8661,
+      "eval_samples_per_second": 25.925,
+      "eval_wer": 0.9416674226617073,
+      "step": 1612
+    },
+    {
+      "epoch": 13.71,
+      "learning_rate": 2.7190278601066983e-05,
+      "loss": 1.6272,
+      "step": 1700
+    },
+    {
+      "epoch": 14.0,
+      "eval_loss": 1.3720481395721436,
+      "eval_runtime": 323.1738,
+      "eval_samples_per_second": 26.06,
+      "eval_wer": 0.9150088775417633,
+      "step": 1736
+    },
+    {
+      "epoch": 14.51,
+      "learning_rate": 2.693623507494284e-05,
+      "loss": 1.5294,
+      "step": 1800
+    },
+    {
+      "epoch": 15.0,
+      "eval_loss": 1.2405216693878174,
+      "eval_runtime": 324.1276,
+      "eval_samples_per_second": 25.984,
+      "eval_wer": 0.8807299024118401,
+      "step": 1860
+    },
+    {
+      "epoch": 15.32,
+      "learning_rate": 2.66821915488187e-05,
+      "loss": 1.4318,
+      "step": 1900
+    },
+    {
+      "epoch": 16.0,
+      "eval_loss": 1.1300945281982422,
+      "eval_runtime": 322.6101,
+      "eval_samples_per_second": 26.106,
+      "eval_wer": 0.8378844234781819,
+      "step": 1984
+    },
+    {
+      "epoch": 16.13,
+      "learning_rate": 2.6428148022694557e-05,
+      "loss": 1.3389,
+      "step": 2000
+    },
+    {
+      "epoch": 16.93,
+      "learning_rate": 2.6174104496570414e-05,
+      "loss": 1.271,
+      "step": 2100
+    },
+    {
+      "epoch": 17.0,
+      "eval_loss": 1.0245275497436523,
+      "eval_runtime": 322.1222,
+      "eval_samples_per_second": 26.145,
+      "eval_wer": 0.7895957802516815,
+      "step": 2108
+    },
+    {
+      "epoch": 17.74,
+      "learning_rate": 2.592006097044627e-05,
+      "loss": 1.1946,
+      "step": 2200
+    },
+    {
+      "epoch": 18.0,
+      "eval_loss": 0.9543349146842957,
+      "eval_runtime": 323.4197,
+      "eval_samples_per_second": 26.04,
+      "eval_wer": 0.7479944531563872,
+      "step": 2232
+    },
+    {
+      "epoch": 18.55,
+      "learning_rate": 2.5666017444322128e-05,
+      "loss": 1.1378,
+      "step": 2300
+    },
+    {
+      "epoch": 19.0,
+      "eval_loss": 0.8853476643562317,
+      "eval_runtime": 323.2454,
+      "eval_samples_per_second": 26.055,
+      "eval_wer": 0.7081427145837923,
+      "step": 2356
+    },
+    {
+      "epoch": 19.35,
+      "learning_rate": 2.5411973918197985e-05,
+      "loss": 1.0823,
+      "step": 2400
+    },
+    {
+      "epoch": 20.0,
+      "eval_loss": 0.8469970226287842,
+      "eval_runtime": 335.6759,
+      "eval_samples_per_second": 25.09,
+      "eval_wer": 0.677920192843535,
+      "step": 2480
+    },
+    {
+      "epoch": 20.16,
+      "learning_rate": 2.5157930392073842e-05,
+      "loss": 1.0365,
+      "step": 2500
+    },
+    {
+      "epoch": 20.96,
+      "learning_rate": 2.4903886865949702e-05,
+      "loss": 0.9959,
+      "step": 2600
+    },
+    {
+      "epoch": 21.0,
+      "eval_loss": 0.8100908994674683,
+      "eval_runtime": 321.969,
+      "eval_samples_per_second": 26.158,
+      "eval_wer": 0.6509765295939659,
+      "step": 2604
+    },
+    {
+      "epoch": 21.77,
+      "learning_rate": 2.4649843339825556e-05,
+      "loss": 0.9687,
+      "step": 2700
+    },
+    {
+      "epoch": 22.0,
+      "eval_loss": 0.7900197505950928,
+      "eval_runtime": 323.9206,
+      "eval_samples_per_second": 26.0,
+      "eval_wer": 0.631653296354376,
+      "step": 2728
+    },
+    {
+      "epoch": 22.58,
+      "learning_rate": 2.4395799813701413e-05,
+      "loss": 0.9366,
+      "step": 2800
+    },
+    {
+      "epoch": 23.0,
+      "eval_loss": 0.7546305060386658,
+      "eval_runtime": 341.914,
+      "eval_samples_per_second": 24.632,
+      "eval_wer": 0.6096732805432796,
+      "step": 2852
+    },
+    {
+      "epoch": 23.39,
+      "learning_rate": 2.4141756287577273e-05,
+      "loss": 0.9048,
+      "step": 2900
+    },
+    {
+      "epoch": 24.0,
+      "eval_loss": 0.7222914099693298,
+      "eval_runtime": 322.3328,
+      "eval_samples_per_second": 26.128,
+      "eval_wer": 0.5865398322987001,
+      "step": 2976
+    },
+    {
+      "epoch": 24.19,
+      "learning_rate": 2.388771276145313e-05,
+      "loss": 0.8833,
+      "step": 3000
+    },
+    {
+      "epoch": 25.0,
+      "learning_rate": 2.3633669235328987e-05,
+      "loss": 0.8561,
+      "step": 3100
+    },
+    {
+      "epoch": 25.0,
+      "eval_loss": 0.6935585141181946,
+      "eval_runtime": 321.6792,
+      "eval_samples_per_second": 26.181,
+      "eval_wer": 0.569432744521196,
+      "step": 3100
+    },
+    {
+      "epoch": 25.8,
+      "learning_rate": 2.3379625709204844e-05,
+      "loss": 0.833,
+      "step": 3200
+    },
+    {
+      "epoch": 26.0,
+      "eval_loss": 0.6709055304527283,
+      "eval_runtime": 392.4001,
+      "eval_samples_per_second": 21.463,
+      "eval_wer": 0.5539586060315445,
+      "step": 3224
+    },
+    {
+      "epoch": 26.61,
+      "learning_rate": 2.31255821830807e-05,
+      "loss": 0.8282,
+      "step": 3300
+    },
+    {
+      "epoch": 27.0,
+      "eval_loss": 0.6760900020599365,
+      "eval_runtime": 350.7656,
+      "eval_samples_per_second": 24.01,
+      "eval_wer": 0.5524552558935213,
+      "step": 3348
+    },
+    {
+      "epoch": 27.42,
+      "learning_rate": 2.287153865695656e-05,
+      "loss": 0.8011,
+      "step": 3400
+    },
+    {
+      "epoch": 28.0,
+      "eval_loss": 0.6472740769386292,
+      "eval_runtime": 323.4247,
+      "eval_samples_per_second": 26.04,
+      "eval_wer": 0.5308510776169308,
+      "step": 3472
+    },
+    {
+      "epoch": 28.23,
+      "learning_rate": 2.2617495130832415e-05,
+      "loss": 0.7905,
+      "step": 3500
+    },
+    {
+      "epoch": 29.0,
+      "eval_loss": 0.6365417838096619,
+      "eval_runtime": 321.7957,
+      "eval_samples_per_second": 26.172,
+      "eval_wer": 0.5260947888181854,
+      "step": 3596
+    },
+    {
+      "epoch": 29.03,
+      "learning_rate": 2.2363451604708276e-05,
+      "loss": 0.7742,
+      "step": 3600
+    },
+    {
+      "epoch": 29.84,
+      "learning_rate": 2.2109408078584133e-05,
+      "loss": 0.7617,
+      "step": 3700
+    },
+    {
+      "epoch": 30.0,
+      "eval_loss": 0.6270097494125366,
+      "eval_runtime": 324.3695,
+      "eval_samples_per_second": 25.964,
+      "eval_wer": 0.5142235066937961,
+      "step": 3720
+    },
+    {
+      "epoch": 30.64,
+      "learning_rate": 2.1855364552459986e-05,
+      "loss": 0.7437,
+      "step": 3800
+    },
+    {
+      "epoch": 31.0,
+      "eval_loss": 0.6166727542877197,
+      "eval_runtime": 379.1426,
+      "eval_samples_per_second": 22.213,
+      "eval_wer": 0.508741462656005,
+      "step": 3844
+    },
+    {
+      "epoch": 31.45,
+      "learning_rate": 2.1601321026335847e-05,
+      "loss": 0.7326,
+      "step": 3900
+    },
+    {
+      "epoch": 32.0,
+      "eval_loss": 0.6096071004867554,
+      "eval_runtime": 320.7646,
+      "eval_samples_per_second": 26.256,
+      "eval_wer": 0.5016394292453441,
+      "step": 3968
+    },
+    {
+      "epoch": 32.26,
+      "learning_rate": 2.1347277500211704e-05,
+      "loss": 0.7259,
+      "step": 4000
+    },
+    {
+      "epoch": 33.0,
+      "eval_loss": 0.594753623008728,
+      "eval_runtime": 340.2561,
+      "eval_samples_per_second": 24.752,
+      "eval_wer": 0.49193245292310883,
+      "step": 4092
+    },
+    {
+      "epoch": 33.06,
+      "learning_rate": 2.109323397408756e-05,
+      "loss": 0.7167,
+      "step": 4100
+    },
+    {
+      "epoch": 33.87,
+      "learning_rate": 2.083919044796342e-05,
+      "loss": 0.706,
+      "step": 4200
+    },
+    {
+      "epoch": 34.0,
+      "eval_loss": 0.5761483907699585,
+      "eval_runtime": 321.498,
+      "eval_samples_per_second": 26.196,
+      "eval_wer": 0.4813571622970153,
+      "step": 4216
+    },
+    {
+      "epoch": 34.68,
+      "learning_rate": 2.0585146921839275e-05,
+      "loss": 0.6949,
+      "step": 4300
+    },
+    {
+      "epoch": 35.0,
+      "eval_loss": 0.5691355466842651,
+      "eval_runtime": 401.8625,
+      "eval_samples_per_second": 20.957,
+      "eval_wer": 0.4750845634452638,
+      "step": 4340
+    },
+    {
+      "epoch": 35.48,
+      "learning_rate": 2.0331103395715132e-05,
+      "loss": 0.6864,
+      "step": 4400
+    },
+    {
+      "epoch": 36.0,
+      "eval_loss": 0.5673930048942566,
+      "eval_runtime": 365.2098,
+      "eval_samples_per_second": 23.061,
+      "eval_wer": 0.4734256943274452,
+      "step": 4464
+    },
+    {
+      "epoch": 36.29,
+      "learning_rate": 2.007705986959099e-05,
+      "loss": 0.6753,
+      "step": 4500
+    },
+    {
+      "epoch": 37.0,
+      "eval_loss": 0.559423565864563,
+      "eval_runtime": 361.5077,
+      "eval_samples_per_second": 23.297,
+      "eval_wer": 0.46351135936548254,
+      "step": 4588
+    },
+    {
+      "epoch": 37.1,
+      "learning_rate": 1.982301634346685e-05,
+      "loss": 0.6758,
+      "step": 4600
+    },
+    {
+      "epoch": 37.9,
+      "learning_rate": 1.9568972817342706e-05,
+      "loss": 0.6551,
+      "step": 4700
+    },
+    {
+      "epoch": 38.0,
+      "eval_loss": 0.5445694327354431,
+      "eval_runtime": 320.4252,
+      "eval_samples_per_second": 26.284,
+      "eval_wer": 0.45804227524267443,
+      "step": 4712
+    },
+    {
+      "epoch": 38.71,
+      "learning_rate": 1.931492929121856e-05,
+      "loss": 0.6579,
+      "step": 4800
+    },
+    {
+      "epoch": 39.0,
+      "eval_loss": 0.5564364790916443,
+      "eval_runtime": 321.0622,
+      "eval_samples_per_second": 26.232,
+      "eval_wer": 0.45992146291520325,
+      "step": 4836
+    },
+    {
+      "epoch": 39.51,
+      "learning_rate": 1.906088576509442e-05,
+      "loss": 0.653,
+      "step": 4900
+    },
+    {
+      "epoch": 40.0,
+      "eval_loss": 0.5391087532043457,
+      "eval_runtime": 321.4848,
+      "eval_samples_per_second": 26.197,
+      "eval_wer": 0.44950169126890527,
+      "step": 4960
+    },
+    {
+      "epoch": 40.32,
+      "learning_rate": 1.8806842238970277e-05,
+      "loss": 0.6409,
+      "step": 5000
+    },
+    {
+      "epoch": 41.0,
+      "eval_loss": 0.5319904685020447,
+      "eval_runtime": 339.4882,
+      "eval_samples_per_second": 24.808,
+      "eval_wer": 0.44666346988763755,
+      "step": 5084
+    },
+    {
+      "epoch": 41.13,
+      "learning_rate": 1.8552798712846134e-05,
+      "loss": 0.6313,
+      "step": 5100
+    },
+    {
+      "epoch": 41.93,
+      "learning_rate": 1.8298755186721995e-05,
+      "loss": 0.6215,
+      "step": 5200
+    },
+    {
+      "epoch": 42.0,
+      "eval_loss": 0.5356938242912292,
+      "eval_runtime": 332.3067,
+      "eval_samples_per_second": 25.344,
+      "eval_wer": 0.44181646168401134,
+      "step": 5208
+    },
+    {
+      "epoch": 42.74,
+      "learning_rate": 1.804471166059785e-05,
+      "loss": 0.6314,
+      "step": 5300
+    },
+    {
+      "epoch": 43.0,
+      "eval_loss": 0.527190625667572,
+      "eval_runtime": 322.6287,
+      "eval_samples_per_second": 26.104,
+      "eval_wer": 0.4371508922901466,
+      "step": 5332
+    },
+    {
+      "epoch": 43.55,
+      "learning_rate": 1.7790668134473705e-05,
+      "loss": 0.6234,
+      "step": 5400
+    },
+    {
+      "epoch": 44.0,
+      "eval_loss": 0.5141627788543701,
+      "eval_runtime": 322.2738,
+      "eval_samples_per_second": 26.133,
+      "eval_wer": 0.4277031142675704,
+      "step": 5456
+    },
+    {
+      "epoch": 44.35,
+      "learning_rate": 1.7536624608349562e-05,
+      "loss": 0.6062,
+      "step": 5500
+    },
+    {
+      "epoch": 45.0,
+      "eval_loss": 0.5133494734764099,
+      "eval_runtime": 471.7285,
+      "eval_samples_per_second": 17.853,
+      "eval_wer": 0.427223597413201,
+      "step": 5580
+    },
+    {
+      "epoch": 45.16,
+      "learning_rate": 1.7282581082225423e-05,
+      "loss": 0.6054,
+      "step": 5600
+    },
+    {
+      "epoch": 45.96,
+      "learning_rate": 1.702853755610128e-05,
+      "loss": 0.6003,
+      "step": 5700
+    },
+    {
+      "epoch": 46.0,
+      "eval_loss": 0.5130398869514465,
+      "eval_runtime": 384.8918,
+      "eval_samples_per_second": 21.881,
+      "eval_wer": 0.42505929161104705,
+      "step": 5704
+    },
+    {
+      "epoch": 46.77,
+      "learning_rate": 1.6774494029977137e-05,
+      "loss": 0.6032,
+      "step": 5800
+    },
+    {
+      "epoch": 47.0,
+      "eval_loss": 0.5147580504417419,
+      "eval_runtime": 323.7827,
+      "eval_samples_per_second": 26.011,
+      "eval_wer": 0.42323194359845,
+      "step": 5828
+    },
+    {
+      "epoch": 47.58,
+      "learning_rate": 1.6520450503852994e-05,
+      "loss": 0.5958,
+      "step": 5900
+    },
+    {
+      "epoch": 48.0,
+      "eval_loss": 0.5133355259895325,
+      "eval_runtime": 372.4972,
+      "eval_samples_per_second": 22.61,
+      "eval_wer": 0.42092507873148355,
+      "step": 5952
+    },
+    {
+      "epoch": 48.39,
+      "learning_rate": 1.626640697772885e-05,
+      "loss": 0.5858,
+      "step": 6000
+    },
+    {
+      "epoch": 49.0,
+      "eval_loss": 0.5059527158737183,
+      "eval_runtime": 322.2789,
+      "eval_samples_per_second": 26.133,
+      "eval_wer": 0.41855341428960224,
+      "step": 6076
+    },
+    {
+      "epoch": 49.19,
+      "learning_rate": 1.6012363451604708e-05,
+      "loss": 0.5856,
+      "step": 6100
+    },
+    {
+      "epoch": 50.0,
+      "learning_rate": 1.5758319925480568e-05,
+      "loss": 0.5772,
+      "step": 6200
+    },
+    {
+      "epoch": 50.0,
+      "eval_loss": 0.49973514676094055,
+      "eval_runtime": 343.25,
+      "eval_samples_per_second": 24.536,
+      "eval_wer": 0.4129806508469304,
+      "step": 6200
+    },
+    {
+      "epoch": 50.8,
+      "learning_rate": 1.5504276399356425e-05,
+      "loss": 0.572,
+      "step": 6300
+    },
+    {
+      "epoch": 51.0,
+      "eval_loss": 0.5019872188568115,
+      "eval_runtime": 322.3073,
+      "eval_samples_per_second": 26.13,
+      "eval_wer": 0.41037570793535594,
+      "step": 6324
+    },
+    {
+      "epoch": 51.61,
+      "learning_rate": 1.525023287323228e-05,
+      "loss": 0.5686,
+      "step": 6400
+    },
+    {
+      "epoch": 52.0,
+      "eval_loss": 0.49171143770217896,
+      "eval_runtime": 383.2843,
+      "eval_samples_per_second": 21.973,
+      "eval_wer": 0.40717460893456536,
+      "step": 6448
+    },
+    {
+      "epoch": 52.42,
+      "learning_rate": 1.4996189347108139e-05,
+      "loss": 0.5669,
+      "step": 6500
+    },
+    {
+      "epoch": 53.0,
+      "eval_loss": 0.49633103609085083,
+      "eval_runtime": 369.1234,
+      "eval_samples_per_second": 22.816,
+      "eval_wer": 0.40889827762729875,
+      "step": 6572
+    },
+    {
+      "epoch": 53.23,
+      "learning_rate": 1.4742145820983994e-05,
+      "loss": 0.5617,
+      "step": 6600
+    },
+    {
+      "epoch": 54.0,
+      "eval_loss": 0.47671106457710266,
+      "eval_runtime": 322.4159,
+      "eval_samples_per_second": 26.122,
+      "eval_wer": 0.4013167273622685,
+      "step": 6696
+    },
+    {
+      "epoch": 54.03,
+      "learning_rate": 1.4488102294859853e-05,
+      "loss": 0.5555,
+      "step": 6700
+    },
+    {
+      "epoch": 54.84,
+      "learning_rate": 1.4234058768735712e-05,
+      "loss": 0.5493,
+      "step": 6800
+    },
+    {
+      "epoch": 55.0,
+      "eval_loss": 0.486770361661911,
+      "eval_runtime": 324.9403,
+      "eval_samples_per_second": 25.919,
+      "eval_wer": 0.40276823784035976,
+      "step": 6820
+    },
+    {
+      "epoch": 55.64,
+      "learning_rate": 1.3980015242611567e-05,
+      "loss": 0.5513,
+      "step": 6900
+    },
+    {
+      "epoch": 56.0,
+      "eval_loss": 0.48170745372772217,
+      "eval_runtime": 322.6243,
+      "eval_samples_per_second": 26.105,
+      "eval_wer": 0.4010316092326434,
+      "step": 6944
+    },
+    {
+      "epoch": 56.45,
+      "learning_rate": 1.3725971716487426e-05,
+      "loss": 0.5414,
+      "step": 7000
+    },
+    {
+      "epoch": 57.0,
+      "eval_loss": 0.47643014788627625,
+      "eval_runtime": 323.5223,
+      "eval_samples_per_second": 26.032,
+      "eval_wer": 0.398711784450694,
+      "step": 7068
+    },
+    {
+      "epoch": 57.26,
+      "learning_rate": 1.3471928190363281e-05,
+      "loss": 0.5442,
+      "step": 7100
+    },
+    {
+      "epoch": 58.0,
+      "eval_loss": 0.47905492782592773,
+      "eval_runtime": 323.3248,
+      "eval_samples_per_second": 26.048,
+      "eval_wer": 0.3988284236855406,
+      "step": 7192
+    },
+    {
+      "epoch": 58.06,
+      "learning_rate": 1.321788466423914e-05,
+      "loss": 0.5456,
+      "step": 7200
+    },
+    {
+      "epoch": 58.87,
+      "learning_rate": 1.2963841138114999e-05,
+      "loss": 0.5347,
+      "step": 7300
+    },
+    {
+      "epoch": 59.0,
+      "eval_loss": 0.4702017605304718,
+      "eval_runtime": 322.1046,
+      "eval_samples_per_second": 26.147,
+      "eval_wer": 0.3954588457899716,
+      "step": 7316
+    },
+    {
+      "epoch": 59.68,
+      "learning_rate": 1.2709797611990854e-05,
+      "loss": 0.536,
+      "step": 7400
+    },
+    {
+      "epoch": 60.0,
+      "eval_loss": 0.4706251621246338,
+      "eval_runtime": 354.4664,
+      "eval_samples_per_second": 23.76,
+      "eval_wer": 0.3925299050038232,
+      "step": 7440
+    },
+    {
+      "epoch": 60.48,
+      "learning_rate": 1.2455754085866713e-05,
+      "loss": 0.5434,
+      "step": 7500
+    },
+    {
+      "epoch": 61.0,
+      "eval_loss": 0.4594361484050751,
+      "eval_runtime": 322.7998,
+      "eval_samples_per_second": 26.09,
+      "eval_wer": 0.38897888829849275,
+      "step": 7564
+    },
+    {
+      "epoch": 61.29,
+      "learning_rate": 1.220171055974257e-05,
+      "loss": 0.5327,
+      "step": 7600
+    },
+    {
+      "epoch": 62.0,
+      "eval_loss": 0.459745854139328,
+      "eval_runtime": 322.7677,
+      "eval_samples_per_second": 26.093,
+      "eval_wer": 0.38711266054094684,
+      "step": 7688
+    },
+    {
+      "epoch": 62.1,
+      "learning_rate": 1.1947667033618427e-05,
+      "loss": 0.5236,
+      "step": 7700
+    },
+    {
+      "epoch": 62.9,
+      "learning_rate": 1.1693623507494285e-05,
+      "loss": 0.5289,
+      "step": 7800
+    },
+    {
+      "epoch": 63.0,
+      "eval_loss": 0.45170098543167114,
+      "eval_runtime": 326.5763,
+      "eval_samples_per_second": 25.789,
+      "eval_wer": 0.3873977786705719,
+      "step": 7812
+    },
+    {
+      "epoch": 63.71,
+      "learning_rate": 1.143957998137014e-05,
+      "loss": 0.5214,
+      "step": 7900
+    },
+    {
+      "epoch": 64.0,
+      "eval_loss": 0.46003150939941406,
+      "eval_runtime": 343.8868,
+      "eval_samples_per_second": 24.491,
+      "eval_wer": 0.3875273778204015,
+      "step": 7936
+    },
+    {
+      "epoch": 64.51,
+      "learning_rate": 1.1185536455246e-05,
+      "loss": 0.521,
+      "step": 8000
+    },
+    {
+      "epoch": 65.0,
+      "eval_loss": 0.4538457691669464,
+      "eval_runtime": 322.9965,
+      "eval_samples_per_second": 26.075,
+      "eval_wer": 0.38370420290042895,
+      "step": 8060
+    },
+    {
+      "epoch": 65.32,
+      "learning_rate": 1.0931492929121856e-05,
+      "loss": 0.5157,
+      "step": 8100
+    },
+    {
+      "epoch": 66.0,
+      "eval_loss": 0.45208266377449036,
+      "eval_runtime": 321.7023,
+      "eval_samples_per_second": 26.179,
+      "eval_wer": 0.3809955806689908,
+      "step": 8184
+    },
+    {
+      "epoch": 66.13,
+      "learning_rate": 1.0677449402997713e-05,
+      "loss": 0.5199,
+      "step": 8200
+    },
+    {
+      "epoch": 66.93,
+      "learning_rate": 1.0423405876873572e-05,
+      "loss": 0.5129,
+      "step": 8300
+    },
+    {
+      "epoch": 67.0,
+      "eval_loss": 0.45266279578208923,
+      "eval_runtime": 321.8831,
+      "eval_samples_per_second": 26.165,
+      "eval_wer": 0.3802309456849963,
+      "step": 8308
+    },
+    {
+      "epoch": 67.74,
+      "learning_rate": 1.0169362350749429e-05,
+      "loss": 0.5121,
+      "step": 8400
+    },
+    {
+      "epoch": 68.0,
+      "eval_loss": 0.4551254212856293,
+      "eval_runtime": 321.9874,
+      "eval_samples_per_second": 26.156,
+      "eval_wer": 0.38033462500486,
+      "step": 8432
+    },
+    {
+      "epoch": 68.55,
+      "learning_rate": 9.915318824625286e-06,
+      "loss": 0.5071,
+      "step": 8500
+    },
+    {
+      "epoch": 69.0,
+      "eval_loss": 0.44843462109565735,
+      "eval_runtime": 322.9859,
+      "eval_samples_per_second": 26.075,
+      "eval_wer": 0.3778333614131491,
+      "step": 8556
+    },
+    {
+      "epoch": 69.35,
+      "learning_rate": 9.661275298501143e-06,
+      "loss": 0.5066,
+      "step": 8600
+    },
+    {
+      "epoch": 70.0,
+      "eval_loss": 0.4638895094394684,
+      "eval_runtime": 321.9358,
+      "eval_samples_per_second": 26.16,
+      "eval_wer": 0.3829525278314174,
+      "step": 8680
+    },
+    {
+      "epoch": 70.16,
+      "learning_rate": 9.407231772377e-06,
+      "loss": 0.515,
+      "step": 8700
+    },
+    {
+      "epoch": 70.96,
+      "learning_rate": 9.153188246252859e-06,
+      "loss": 0.5058,
+      "step": 8800
+    },
+    {
+      "epoch": 71.0,
+      "eval_loss": 0.45568934082984924,
+      "eval_runtime": 400.5646,
+      "eval_samples_per_second": 21.025,
+      "eval_wer": 0.37642073068000675,
+      "step": 8804
+    },
+    {
+      "epoch": 71.77,
+      "learning_rate": 8.899144720128716e-06,
+      "loss": 0.4984,
+      "step": 8900
+    },
+    {
+      "epoch": 72.0,
+      "eval_loss": 0.4499472975730896,
+      "eval_runtime": 396.1535,
+      "eval_samples_per_second": 21.259,
+      "eval_wer": 0.37675768846956365,
+      "step": 8928
+    },
+    {
+      "epoch": 72.58,
+      "learning_rate": 8.645101194004573e-06,
+      "loss": 0.4986,
+      "step": 9000
+    },
+    {
+      "epoch": 73.0,
+      "eval_loss": 0.4449363946914673,
+      "eval_runtime": 613.581,
+      "eval_samples_per_second": 13.726,
+      "eval_wer": 0.37433418436775057,
+      "step": 9052
+    },
+    {
+      "epoch": 73.39,
+      "learning_rate": 8.39105766788043e-06,
+      "loss": 0.5002,
+      "step": 9100
+    },
+    {
+      "epoch": 74.0,
+      "eval_loss": 0.44110018014907837,
+      "eval_runtime": 394.6941,
+      "eval_samples_per_second": 21.338,
+      "eval_wer": 0.37380282785344926,
+      "step": 9176
+    },
+    {
+      "epoch": 74.19,
+      "learning_rate": 8.137014141756288e-06,
+      "loss": 0.4913,
+      "step": 9200
+    },
+    {
+      "epoch": 75.0,
+      "learning_rate": 7.882970615632145e-06,
+      "loss": 0.498,
+      "step": 9300
+    },
+    {
+      "epoch": 75.0,
+      "eval_loss": 0.4465697705745697,
+      "eval_runtime": 319.8787,
+      "eval_samples_per_second": 26.329,
+      "eval_wer": 0.3730900325293866,
+      "step": 9300
+    },
+    {
+      "epoch": 75.8,
+      "learning_rate": 7.628927089508002e-06,
+      "loss": 0.491,
+      "step": 9400
+    },
+    {
+      "epoch": 76.0,
+      "eval_loss": 0.4530308246612549,
+      "eval_runtime": 320.8191,
+      "eval_samples_per_second": 26.252,
+      "eval_wer": 0.37500809994686435,
+      "step": 9424
+    },
+    {
+      "epoch": 76.61,
+      "learning_rate": 7.37488356338386e-06,
+      "loss": 0.4861,
+      "step": 9500
+    },
+    {
+      "epoch": 77.0,
+      "eval_loss": 0.4457537829875946,
+      "eval_runtime": 486.5829,
+      "eval_samples_per_second": 17.308,
+      "eval_wer": 0.3743860240276824,
+      "step": 9548
+    },
+    {
+      "epoch": 77.42,
+      "learning_rate": 7.120840037259717e-06,
+      "loss": 0.4931,
+      "step": 9600
+    },
+    {
+      "epoch": 78.0,
+      "eval_loss": 0.43678486347198486,
+      "eval_runtime": 322.4494,
+      "eval_samples_per_second": 26.119,
+      "eval_wer": 0.3692538976944311,
+      "step": 9672
+    },
+    {
+      "epoch": 78.23,
+      "learning_rate": 6.866796511135574e-06,
+      "loss": 0.4918,
+      "step": 9700
+    },
+    {
+      "epoch": 79.0,
+      "eval_loss": 0.4429604113101959,
+      "eval_runtime": 323.2104,
+      "eval_samples_per_second": 26.057,
+      "eval_wer": 0.37231243763040917,
+      "step": 9796
+    },
+    {
+      "epoch": 79.03,
+      "learning_rate": 6.612752985011432e-06,
+      "loss": 0.4884,
+      "step": 9800
+    },
+    {
+      "epoch": 79.84,
+      "learning_rate": 6.35870945888729e-06,
+      "loss": 0.4851,
+      "step": 9900
+    },
+    {
+      "epoch": 80.0,
+      "eval_loss": 0.44754448533058167,
+      "eval_runtime": 321.4015,
+      "eval_samples_per_second": 26.204,
+      "eval_wer": 0.36930573735436295,
+      "step": 9920
+    },
+    {
+      "epoch": 80.64,
+      "learning_rate": 6.104665932763147e-06,
+      "loss": 0.4857,
+      "step": 10000
+    },
+    {
+      "epoch": 81.0,
+      "eval_loss": 0.4332858920097351,
+      "eval_runtime": 321.8635,
+      "eval_samples_per_second": 26.166,
+      "eval_wer": 0.36658415520794185,
+      "step": 10044
+    },
+    {
+      "epoch": 81.45,
+      "learning_rate": 5.850622406639005e-06,
+      "loss": 0.4864,
+      "step": 10100
+    },
+    {
+      "epoch": 82.0,
+      "eval_loss": 0.4359518885612488,
+      "eval_runtime": 321.9658,
+      "eval_samples_per_second": 26.158,
+      "eval_wer": 0.36487344643019143,
+      "step": 10168
+    },
+    {
+      "epoch": 82.26,
+      "learning_rate": 5.596578880514862e-06,
+      "loss": 0.4851,
+      "step": 10200
+    },
+    {
+      "epoch": 83.0,
+      "eval_loss": 0.43162801861763,
+      "eval_runtime": 357.942,
+      "eval_samples_per_second": 23.529,
+      "eval_wer": 0.36326641697230466,
+      "step": 10292
+    },
+    {
+      "epoch": 83.06,
+      "learning_rate": 5.342535354390719e-06,
+      "loss": 0.4808,
+      "step": 10300
+    },
+    {
+      "epoch": 83.87,
+      "learning_rate": 5.088491828266577e-06,
+      "loss": 0.4859,
+      "step": 10400
+    },
+    {
+      "epoch": 84.0,
+      "eval_loss": 0.4436122179031372,
+      "eval_runtime": 401.5799,
+      "eval_samples_per_second": 20.972,
+      "eval_wer": 0.36780238721633984,
+      "step": 10416
+    },
+    {
+      "epoch": 84.68,
+      "learning_rate": 4.8344483021424344e-06,
+      "loss": 0.4777,
+      "step": 10500
+    },
+    {
+      "epoch": 85.0,
+      "eval_loss": 0.4375947117805481,
+      "eval_runtime": 347.8526,
+      "eval_samples_per_second": 24.211,
+      "eval_wer": 0.368061585515999,
+      "step": 10540
+    },
+    {
+      "epoch": 85.48,
+      "learning_rate": 4.5804047760182914e-06,
+      "loss": 0.4816,
+      "step": 10600
+    },
+    {
+      "epoch": 86.0,
+      "eval_loss": 0.43457648158073425,
+      "eval_runtime": 364.1744,
+      "eval_samples_per_second": 23.126,
+      "eval_wer": 0.36431617008592426,
+      "step": 10664
+    },
+    {
+      "epoch": 86.29,
+      "learning_rate": 4.3263612498941484e-06,
+      "loss": 0.4771,
+      "step": 10700
+    },
+    {
+      "epoch": 87.0,
+      "eval_loss": 0.4322541058063507,
+      "eval_runtime": 376.0177,
+      "eval_samples_per_second": 22.398,
+      "eval_wer": 0.36263138113813975,
+      "step": 10788
+    },
+    {
+      "epoch": 87.1,
+      "learning_rate": 4.074858159031247e-06,
+      "loss": 0.4691,
+      "step": 10800
+    },
+    {
+      "epoch": 87.9,
+      "learning_rate": 3.820814632907104e-06,
+      "loss": 0.4767,
+      "step": 10900
+    },
+    {
+      "epoch": 88.0,
+      "eval_loss": 0.43458959460258484,
+      "eval_runtime": 322.6421,
+      "eval_samples_per_second": 26.103,
+      "eval_wer": 0.3655473620093052,
+      "step": 10912
+    },
+    {
+      "epoch": 88.71,
+      "learning_rate": 3.566771106782962e-06,
+      "loss": 0.4751,
+      "step": 11000
+    },
+    {
+      "epoch": 89.0,
+      "eval_loss": 0.436272531747818,
+      "eval_runtime": 323.9622,
+      "eval_samples_per_second": 25.997,
+      "eval_wer": 0.36510672489988466,
+      "step": 11036
+    },
+    {
+      "epoch": 89.51,
+      "learning_rate": 3.31272758065882e-06,
+      "loss": 0.4818,
+      "step": 11100
+    },
+    {
+      "epoch": 90.0,
+      "eval_loss": 0.43062761425971985,
+      "eval_runtime": 322.2604,
+      "eval_samples_per_second": 26.134,
+      "eval_wer": 0.36658415520794185,
+      "step": 11160
+    },
+    {
+      "epoch": 90.32,
+      "learning_rate": 3.058684054534677e-06,
+      "loss": 0.4788,
+      "step": 11200
+    },
+    {
+      "epoch": 91.0,
+      "eval_loss": 0.4332570731639862,
+      "eval_runtime": 323.8045,
+      "eval_samples_per_second": 26.01,
+      "eval_wer": 0.36097251202032116,
+      "step": 11284
+    },
+    {
+      "epoch": 91.13,
+      "learning_rate": 2.8046405284105347e-06,
+      "loss": 0.4803,
+      "step": 11300
+    },
+    {
+      "epoch": 91.93,
+      "learning_rate": 2.5505970022863917e-06,
+      "loss": 0.4729,
+      "step": 11400
+    },
+    {
+      "epoch": 92.0,
+      "eval_loss": 0.4346713721752167,
+      "eval_runtime": 323.6128,
+      "eval_samples_per_second": 26.025,
+      "eval_wer": 0.3635515351019297,
+      "step": 11408
+    },
+    {
+      "epoch": 92.74,
+      "learning_rate": 2.296553476162249e-06,
+      "loss": 0.4751,
+      "step": 11500
+    },
+    {
+      "epoch": 93.0,
+      "eval_loss": 0.43105602264404297,
+      "eval_runtime": 322.0335,
+      "eval_samples_per_second": 26.153,
+      "eval_wer": 0.36287761952281594,
+      "step": 11532
+    },
+    {
+      "epoch": 93.55,
+      "learning_rate": 2.0425099500381066e-06,
+      "loss": 0.4735,
+      "step": 11600
+    },
+    {
+      "epoch": 94.0,
+      "eval_loss": 0.4452572464942932,
+      "eval_runtime": 322.6054,
+      "eval_samples_per_second": 26.106,
+      "eval_wer": 0.36640271639818045,
+      "step": 11656
+    },
+    {
+      "epoch": 94.35,
+      "learning_rate": 1.788466423913964e-06,
+      "loss": 0.4676,
+      "step": 11700
+    },
+    {
+      "epoch": 95.0,
+      "eval_loss": 0.43202486634254456,
+      "eval_runtime": 322.7019,
+      "eval_samples_per_second": 26.098,
+      "eval_wer": 0.3624110625834295,
+      "step": 11780
+    },
+    {
+      "epoch": 95.16,
+      "learning_rate": 1.5344228977898214e-06,
+      "loss": 0.4754,
+      "step": 11800
+    },
+    {
+      "epoch": 95.96,
+      "learning_rate": 1.2803793716656788e-06,
+      "loss": 0.467,
+      "step": 11900
+    },
+    {
+      "epoch": 96.0,
+      "eval_loss": 0.4379962384700775,
+      "eval_runtime": 322.1888,
+      "eval_samples_per_second": 26.14,
+      "eval_wer": 0.36453648864063454,
+      "step": 11904
+    },
+    {
+      "epoch": 96.77,
+      "learning_rate": 1.0263358455415363e-06,
+      "loss": 0.4796,
+      "step": 12000
+    },
+    {
+      "epoch": 97.0,
+      "eval_loss": 0.43832120299339294,
+      "eval_runtime": 320.8953,
+      "eval_samples_per_second": 26.245,
+      "eval_wer": 0.3636552144217934,
+      "step": 12028
+    },
+    {
+      "epoch": 97.58,
+      "learning_rate": 7.722923194173936e-07,
+      "loss": 0.4731,
+      "step": 12100
+    },
+    {
+      "epoch": 98.0,
+      "eval_loss": 0.43025869131088257,
+      "eval_runtime": 323.1419,
+      "eval_samples_per_second": 26.063,
+      "eval_wer": 0.36028563652622436,
+      "step": 12152
+    },
+    {
+      "epoch": 98.39,
+      "learning_rate": 5.18248793293251e-07,
+      "loss": 0.4666,
+      "step": 12200
+    },
+    {
+      "epoch": 99.0,
+      "eval_loss": 0.43907731771469116,
+      "eval_runtime": 321.8174,
+      "eval_samples_per_second": 26.17,
+      "eval_wer": 0.36353857518694677,
+      "step": 12276
+    },
+    {
+      "epoch": 99.19,
+      "learning_rate": 2.642052671691083e-07,
+      "loss": 0.476,
+      "step": 12300
+    },
+    {
+      "epoch": 100.0,
+      "learning_rate": 1.0161741044965706e-08,
+      "loss": 0.4705,
+      "step": 12400
+    },
+    {
+      "epoch": 100.0,
+      "eval_loss": 0.43463364243507385,
+      "eval_runtime": 323.3381,
+      "eval_samples_per_second": 26.047,
+      "eval_wer": 0.36169826725936677,
+      "step": 12400
+    },
+    {
+      "epoch": 100.0,
+      "step": 12400,
+      "total_flos": 0,
+      "train_runtime": 215429.416,
+      "train_samples_per_second": 0.058
+    }
+  ],
+  "max_steps": 12400,
+  "num_train_epochs": 100,
+  "total_flos": 0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63f05ebddc0741251fa05cbd314be2f82a6628b6d155e7f3d1b8be8546e05e09
+size 2543

vocab.json

@@ -0,0 +1 @@
+{"<pad>": 0, "<s>": 1, "</s>": 2, "<unk>": 3, "|": 4, "-": 5, "ё": 6, "а": 7, "б": 8, "в": 9, "г": 10, "д": 11, "е": 12, "ж": 13, "з": 14, "и": 15, "й": 16, "к": 17, "л": 18, "м": 19, "н": 20, "о": 21, "п": 22, "р": 23, "с": 24, "т": 25, "у": 26, "ф": 27, "х": 28, "ц": 29, "ч": 30, "ш": 31, "щ": 32, "ъ": 33, "ы": 34, "ь": 35, "э": 36, "ю": 37, "я": 38}