ground-ckpts

Final training checkpoints for grounded segmentation / referring-expression work built on Qwen3-VL, plus MM-GroundingDINO baseline finetunes.

This is a research checkpoint archive, not a packaged model release. Each run contributes only its last checkpoint, stored in native training format (torch.save state dicts) with optimizer state, so a run can be resumed as-is.

Layout

<group>/<run>/step<N>/
├── model.pt       # model state dict       (~1.7 GB)
├── optimizer.pt   # optimizer state        (~3.4 GB)
└── train/rank*.pt # per-rank trainer state (~18 KB each)

MM-GroundingDINO baselines follow the mmdetection convention instead: the final epoch_15.pth plus its training config .py and last_checkpoint.

Contents

85 runs, one final checkpoint each, plus 3 baselines — ~420 GB total.

group	runs	note
qwen3_5/full_v0	14	early full-model runs
qwen3_5/full_v1	28	weight_compare loss-weight sweep
qwen3_5/ablation_v3	27	single-design-choice ablation sweep
qwen3_5/sa1b_pretrain	2	SA-1B pretraining (backbone frozen / tuned)
qwen3_5/video_v0	4	video tube-assignment extension
qwen3_5/video_v1	4	video, revised assignment strategies
qwen3_5/video_v1_actions	3	+ temporal action tokens
ft_qwen3_5/full_v4, ft_qwen3_5/release	2	finetuning runs
baselines/mm_gdino{,_swinl,_swinl_origin}	3	MM-GroundingDINO finetunes

Sizes: 149.8 GB of model.pt, 289.9 GB of optimizer.pt, 10.4 GB of baseline .pth.

Downloading

Fetch only what you need — weights alone, without optimizer state:

from huggingface_hub import snapshot_download

snapshot_download(
    "royguw/ground-ckpts",
    allow_patterns=["ft_qwen3_5/release/**/model.pt"],
    local_dir="ckpts",
)