This is a full-parameter supervised fine-tuning baseline for research and evaluation of Android visual-language agents. It is not intended for safety-critical, autonomous production use.

The remaining pinned hyperparameters are:

{ "acc_strategy": "token", "adam_beta1": 0.9, "adam_beta2": 0.95, "adam_epsilon": 1e-08, "average_tokens_across_devices": true, "bf16": true, "data_seed": 42, "dataloader_num_workers": 0, "dataset_num_proc": 1, "deepspeed_offload": false, "deepspeed_stage": 2, "fp16": false, "freeze_aligner": true, "freeze_llm": false, "freeze_parameters": [ "model.visual", "model.visual.merger" ], "freeze_vit": true, "global_world_size": 4, "gradient_accumulation_steps": 4, "gradient_checkpointing": true, "lazy_tokenize": true, "learning_rate": 5e-07, "load_best_model_at_end": false, "logging_steps": 5, "loss_scale": "default", "lr_scheduler_type": "cosine", "max_grad_norm": 1.0, "max_length": 8192, "max_pixels": 2629536, "max_steps": -1, "num_train_epochs": 1.0, "optim": "adamw_torch", "packing": false, "padding_free": false, "padding_side": "right", "per_device_eval_batch_size": 1, "per_device_train_batch_size": 2, "remove_unused_columns": true, "save_only_model": true, "save_strategy": "epoch", "save_total_limit": 1, "seed": 42, "split_dataset_ratio": 0.0, "template": "qwen2_5_vl", "torch_dtype": "bfloat16", "train_dataloader_shuffle": true, "train_type": "full", "truncation_strategy": "delete", "use_hf": true, "use_liger_kernel": false, "warmup_ratio": 0.05, "weight_decay": 0.1 }

This checkpoint is intended to measure base-SFT transfer from real-world AMEX navigation data and to support controlled held-out and AndroidWorld research evaluation. Results should report the immutable data/base revisions above and should not be interpreted as a general safety or reliability guarantee.