mms-khotbah-lora

This model is a fine-tuned version of facebook/mms-1b-all on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.4532
• Wer: 0.5076

Model description

More information needed

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.9330	0.0889	50	2.3807	0.7022
8.9734	0.1778	100	1.7969	0.6482
7.4418	0.2667	150	1.6738	0.5730
6.9816	0.3556	200	1.5701	0.5564
6.6540	0.4444	250	1.5545	0.5439
6.9938	0.5333	300	1.5175	0.5359
6.7533	0.6222	350	1.5268	0.5277
7.4097	0.7111	400	1.4680	0.5309
7.6745	0.8	450	1.4981	0.5179
6.8624	0.8889	500	1.4933	0.5154
6.4320	0.9778	550	1.4819	0.5125
6.2416	1.0658	600	1.4602	0.5152
5.6667	1.1547	650	1.4874	0.5079
6.6122	1.2436	700	1.4509	0.5106
5.7469	1.3324	750	1.5019	0.5068
6.5593	1.4213	800	1.4524	0.5062
6.6418	1.5102	850	1.4339	0.5106
6.5021	1.5991	900	1.4704	0.5016
6.7905	1.688	950	1.4382	0.5043
6.7729	1.7769	1000	1.4342	0.5060
5.2900	1.8658	1050	1.4532	0.5076

Framework versions

• PEFT 0.18.1
• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.8.3
• Tokenizers 0.22.2