ViOcean/qwen3vl-vln-new-weights

Qwen3-VL VLN Completed Weights

This folder is a Hugging Face upload staging area for completed weights from /home/zhangshan/syp/myvln/qwen3VL_vln.

The files were staged with hardlinks from outputs/, so they appear as regular files but do not duplicate local disk usage.

Included Final Models

Subdir	Role	R2R val_unseen result
final_models/qwen3vl_streamvln_r2r_navida_random_stop2_zero3	NAVIDA-style R2R baseline	SR 0.4252 / SPL 0.3818 / NE 6.5368 under the current best-policy eval
final_models/qwen3vl_streamvln_r2r_depth_stage2_current_patch2_w05to01_hold30_zero3	Best completed R2R-only depth auxiliary model, current-frame depth, patch=2, dynamic depth weight 0.5->0.1 with 30% hold	SR 0.4584 / SPL 0.4104 / NE 6.4373 under the current best-policy eval

Included Depth-Head Warmup Artifacts

Subdir	Role
depth_heads/qwen3vl_streamvln_r2r_depth_stage1_current_patch2_zero3	R2R current-frame depth head warmup used before the best R2R-only Stage2 run
depth_heads/qwen3vl_streamvln_r2r_depth_stage1_zero3	Older R2R Stage1 depth head artifact kept for reproducibility
depth_heads/qwen3vl_streamvln_r2r_rxr_depth_stage1_current_patch2_zero3	R2R+RxR current-frame depth head warmup

Explicitly Excluded

The following model is not staged because its R2R val_unseen evaluation is still running:

outputs/qwen3vl_streamvln_r2r_rxr_depth_stage2_current_patch2_w05to01_hold30_zero3

Add it only after final_metrics.json is produced and the result is checked.

Upload

Example:

huggingface-cli login
huggingface-cli upload <org-or-user>/<repo-name> hf_upload_weights_completed .

If using Git directly, keep .safetensors files in Git LFS.