vi_whisper-medium

This model is a one shot fine-tuned version of openai/whisper-medium on the Vin100h dataset. It achieves the following results on the evaluation set: - Loss: 0.2894 - Wer: 21.68 on whisper-quantized model by CTranslate2

Model description

To use quantized model , firstly , read the doc of how to use CTranslate2 converter and Faster Whisper repo in here:

• CTranslate2
• Faster-Whipser

Intended uses & limitations

More information needed

Training and evaluation data

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
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• training_steps: 8000

Framework versions

• Transformers 4.32.0.dev0
• Pytorch 2.0.1+cu11.8
• Datasets 2.13.1
• Tokenizers 0.13.3
• PEFT 0.5.0.dev0

Training procedure

The following bitsandbytes quantization config was used during training:

• load_in_8bit: True
• load_in_4bit: False
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: fp4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float32