This model is a fine-tuned version of canopylabs/orpheus-3b-0.1-pretrained on 'SUST-CSE-Speech/banspeech' dataset. Training was ran for 10 epochs on an A100PCE. It achieves the following results on the evaluation set:

Loss: 1.0226

The training scripts were forked from the original repo to include eval steps. Fork can be cloned from: Github DISCLAIMER: This model is only to signify the functioning of the training scripts for SFT, as most lean towards the Unsloth alternative, model will be improved on more datasets for better results on Bangla TTS.

Model Details

Model Capabilities

Human-Like Speech: Minimally fine-tune to produce natural intonation, emotion, and rhythm that is superior to SOTA closed source models
Zero-Shot Voice Cloning: Clone voices without prior fine-tuning

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
optimizer: ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss
0.9649	0.3092	2000	1.2818
0.5303	0.6184	4000	1.1270
0.9436	0.9276	6000	1.0226

Framework versions

Transformers 4.55.1
Pytorch 2.7.1+cu126
Datasets 4.0.0
Tokenizers 0.21.4

noor-raghib-12/orpheus_3b_sft_fp32

orpheus_3b_sft_fp32