9b-41-2

<!-- 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. -->

qwen35_caption_galore

This model is a fine-tuned version of /workspace/models/Qwen3.5-9B on the my_caption dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

same as 9b-41, freeze lm_head + embed, every other step freeze bottom 16 lang

Training hyperparameters

The following hyperparameters were used during training:

• family_to_muon_lr = { "language": _fallback(getattr(training_args, "language_muon_lr", 3e-5), language_lr), "vision": _fallback(getattr(training_args, "vision_muon_lr", 3e-5), vision_lr), "merger": _fallback(getattr(training_args, "merger_muon_lr", 3e-5), merger_lr), }

  family_to_adamw_lr = { "language": _fallback(getattr(training_args, "language_adamw_lr", 3e-5), language_lr), "vision": _fallback(getattr(training_args, "vision_adamw_lr", 3e-6), vision_lr), "merger": _fallback(getattr(training_args, "merger_adamw_lr", 3e-5), merger_lr), }

• train_batch_size: 2

• eval_batch_size: 8

• seed: 42

• distributed_type: multi-GPU

• gradient_accumulation_steps: 32

• total_train_batch_size: 64

• 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: 3

Training results

Framework versions

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