Kira-Floris/TranslateGemma-4B-RW2EN

sft_stage1_rw_en

This model is a fine-tuned version of google/translategemma-4b-it on the synthetic_rw_en_gemma__train and the expert_rw_en_gemma__train datasets. It achieves the following results on the evaluation set:

• Loss: 0.7176

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: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 7
• gradient_accumulation_steps: 8
• total_train_batch_size: 56
• total_eval_batch_size: 7
• 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: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.0598	0.0570	500	1.0485
0.9572	0.1139	1000	0.9877
0.952	0.1709	1500	0.9316
0.9124	0.2278	2000	0.8920
0.8772	0.2848	2500	0.8695
0.8855	0.3418	3000	0.8423
0.7964	0.3987	3500	0.8256
0.8643	0.4557	4000	0.8093
0.7972	0.5126	4500	0.7886
0.7519	0.5696	5000	0.7725
0.7513	0.6266	5500	0.7591
0.7877	0.6835	6000	0.7449
0.7322	0.7405	6500	0.7371
0.7346	0.7974	7000	0.7274
0.7603	0.8544	7500	0.7215
0.7201	0.9114	8000	0.7186
0.7025	0.9683	8500	0.7174

Framework versions

• Transformers 4.57.1
• Pytorch 2.9.1+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2