metadata

language:
  - la
license: mit
tags:
  - generated_from_trainer
datasets:
  - thiagolira/LatinYoutube
metrics:
  - wer
base_model: facebook/w2v-bert-2.0
model-index:
  - name: CiceroASR
    results: []

CiceroASR

This model is a fine-tuned version of facebook/w2v-bert-2.0 for the transcription of Classical Latin!

Example from the Aeneid: Transcription: arma virumque cano (Of arms and men I sing)

Example from Genesis: Transcription (little error there): creavit deus chaelum et terram (In the beggining God created the heaven and the earth)

It achieves the following results on the evaluation set of my dataset Latin Youtube:

Loss: 0.5026
Wer: 0.1651

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 300
num_epochs: 15
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9864	1.14	50	2.4639	1.0
0.7134	2.27	100	0.4891	0.3601
0.5196	3.41	150	0.5267	0.3022
0.3779	4.55	200	0.4407	0.2369
0.3818	5.68	250	0.4516	0.2360
0.3	6.82	300	0.4365	0.2379
0.3252	7.95	350	0.4238	0.2183
0.2736	9.09	400	0.4609	0.2034
0.1588	10.23	450	0.4007	0.2239
0.1223	11.36	500	0.4892	0.1987
0.0859	12.5	550	0.5393	0.1772
0.0575	13.64	600	0.4629	0.1744
0.0464	14.77	650	0.5026	0.1651

Framework versions

Transformers 4.38.0
Pytorch 2.1.0+cu121
Datasets 2.17.1
Tokenizers 0.15.2