qwen35_caption_galore

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• family_to_muon_lr = { "language": _fallback(getattr(training_args, "language_muon_lr", 1e-1), language_lr), "vision": _fallback(getattr(training_args, "vision_muon_lr", 2e-5), vision_lr), "merger": _fallback(getattr(training_args, "merger_muon_lr", 1e-1), merger_lr), }

  0.95 adamw beta2, 0.02 merger wd, language and merger adamw, vision muon

  family_to_adamw_lr = { "language": _fallback(getattr(training_args, "language_adamw_lr", 3e-6), language_lr), "vision": _fallback(getattr(training_args, "vision_adamw_lr", 4e-6), vision_lr), "merger": _fallback(getattr(training_args, "merger_adamw_lr", 1.5e-5), merger_lr), }
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• 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: cosine_with_min_lr
• lr_scheduler_warmup_steps: 0.05
• num_epochs: 4

Training results

Framework versions

• Transformers 5.5.3
• Pytorch 2.11.0+cu130
• Datasets 4.0.0
• Tokenizers 0.22.2