amitabh3/revisit-verifier-lora

Revisit verifier adapters

LoRA adapters that turn an 8B vision-language model into the change verifier of Revisit. Given two crops of a proposed change, a side-by-side context image and the 3D/appearance scores of the proposal, the model answers moved, added, removed or none with a confidence. One subfolder per adapter:

Subfolder	Base model	RL	LoRA rank	Steps	H100 hours	Used for
qwen3vl8b-gdpo-rlcr-step2460	Qwen3-VL-8B-Instruct	GDPO + RLCR	32	2,460	755	room pipeline
sensenova8b-gdpo-rlcr-step2290	SenseNova-SI-1.3-Qwen3-VL-8B	GDPO + RLCR	32	2,290	705	room pipeline
qwen3vl8b-grpo-step45	Qwen3-VL-8B-Instruct	GRPO	16	45	~1	SceneDiff table
qwen3vl8b-gdpo-step45	Qwen3-VL-8B-Instruct	GDPO	16	45	~1	SceneDiff table
qwen3vl8b-gdpo-rlcr-step45	Qwen3-VL-8B-Instruct	GDPO + RLCR	16	45	~1	SceneDiff table

H100 hours cover RL training only: the room adapters ran around ~14-16 hours on six 8-H100 nodes, the step-45 adapters about 8 minutes on one 8-GPU machine.

revisit.json in each subfolder records the base model and RL variant; Revisit reads it to pick the matching fusion rule.

Training data

Evidence cards built from SceneDiff validation scenes, labelled by matching each proposal to SceneDiff's ground truth. The test split was never used for training. The room adapters pool five Stage-1 variants; one of them (Pi3X + DINOv3 + SAM 3) was produced with a bug that placed SAM 3 point prompts outside the image, so its masks are coarse and most of its cards are labelled "none". Details: docs/training.md.

Use

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3-VL-8B-Instruct", torch_dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(base, "amitabh3/revisit-verifier-lora", subfolder="qwen3vl8b-gdpo-rlcr-step2460")

The prompt format and the fusion with Stage-1 scores live in revisit/verifier.py; the adapters are meant to be used through Revisit.

License

Apache-2.0, the license of both base models.