RelMLLM-8B

RelMLLM-8B is the recommended merged checkpoint accompanying Beyond Paired Inputs: Learning to Leverage RGB-T Cross-Modal Relations in Multimodal Large Language Models. It is derived from Qwen3-VL-8B-Instruct and accepts aligned RGB and thermal-infrared observations.

Use the inference code in the companion RelMLLM repository. The RGB image must be provided first and the TIR image second.

Selected results

Benchmark Metric Score
RGBT-GroundBench Acc@0.5 73.1628
Degraded-1K Acc@0.5 59.80
RGB-Th-Bench macro QAcc 75.6745
DV-1K-Balanced Accuracy 78.40

These are complete archived evaluations of the recovered checkpoint, not exact paper-table reproduction. Paper GroundBench/Degraded scores were 73.3011/61.10. Code, prompts, dataset downloads and custom-set reconstruction limitations: RelMLLM GitHub.

The model is intended for research, not safety-critical sensor-health or surveillance decisions. Reliability fields are task-conditioned evidence utilities rather than physical sensor diagnoses. See the companion repository for protocols, limitations, code, license, and attribution.

Downloads last month
16
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tyy1850819840/RelMLLM-8B

Finetuned
(755)
this model