ForgeLoc-R1 checkpoints

Qwen2.5-VL-7B-Instruct fine-tuned with GRPO to localize the manipulated segments of a video. Every folder holds a full checkpoint in Hugging Face format.

Folder	Model	Training data	Test mIoU (%)
af	ForgeLoc-R1	ActivityForensics	64.27
ddl	ForgeLoc-R1	DDL	53.03
baselines/af_grpo	GRPO	ActivityForensics	51.28
baselines/af_video-r1	Video-R1	ActivityForensics	54.29
baselines/af_tempsamp-r1	TempSamp-R1	ActivityForensics	59.93
baselines/ddl_grpo	GRPO	DDL	31.89
baselines/ddl_video-r1	Video-R1	DDL	33.25
baselines/ddl_tempsamp-r1	TempSamp-R1	DDL	50.92
ablations/af_noanchor	without anchor injection	ActivityForensics	58.61
ablations/af_noprune	without advantage pruning	ActivityForensics	62.05
ablations/af_noos	without count balancing (m=1)	ActivityForensics	59.53
ablations/af_notransform	without the asymmetric advantage transform	ActivityForensics	62.90
ablations/af_bestoverlap	set term replaced by the best-overlap soft F1	ActivityForensics	65.35
ablations/af_g16	G'=16	ActivityForensics	63.53
ablations/af_eps0.01	Sinkhorn coefficient 0.01	ActivityForensics	63.49
ablations/af_seed7	ForgeLoc-R1, seed 7	ActivityForensics	63.27
ablations/ddl_noos	without count balancing (m=1)	DDL	50.17

mIoU is measured on the test splits: 1,565 ActivityForensics videos and 1,500 DDL videos.

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "sdzt/ForgeLoc-R1", subfolder="af", torch_dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("sdzt/ForgeLoc-R1", subfolder="af")