magnolia-resin/whisper-small-sjd-v6

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

Whisper Small Sjd - Rus Base - V6

This model is a fine-tuned version of openai/whisper-small on the Kildin Sami V2 Dataset. It achieves the following results on the evaluation set:

• —Loss: 1.2228
• —Wer: 68.5524

Model description

This model set: processor = WhisperProcessor.frompretrained("openai/whisper-small", language="Russian", task="transcribe") and generationconfig.language = "Russian"

Intended uses & limitations

MA Thesis project, intended to test automatic fieldwork transcription process.

Training and evaluation data

80:20 Split Training:Evaluation.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 1e-05
• —trainbatchsize: 16
• —evalbatchsize: 16
• —seed: 42
• —optimizer: Use OptimizerNames.ADAMWTORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
• —lrschedulertype: constantwithwarmup
• —lrschedulerwarmup_steps: 50
• —training_steps: 500
• —mixedprecisiontraining: Native AMP

Training results

Framework versions

• —Transformers 4.51.0.dev0
• —Pytorch 2.6.0+cu124
• —Datasets 3.4.2.dev0
• —Tokenizers 0.21.1