README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
library_name: peft
pipeline_tag: video-text-to-text
tags:
  - video-temporal-grounding
  - moment-retrieval
  - qwen2.5-vl
  - neurips-2026
---

# EVIDENT

Checkpoints of **EVIDENT: Routing MLLM Adaptation through Entity-Grounded Visual Evidence for Cross-Domain Video Temporal Grounding** (NeurIPS 2026).

[[Paper]](https://arxiv.org/abs/2605.26104) [[Code]](https://github.com/KHU-VLL/EVIDENT)

Each folder holds an EVIDENT model fine-tuned from [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) on one source dataset: LoRA weights (`adapter_model.safetensors`), EB Adapter weights (`eb_adapter.safetensors`) and the adapter/prompt configuration (`evident_config.json`).

| Folder | Source | Frames | Timestamps | λ (EBD) |
|---|---|:-:|---|:-:|
| [`charades/`](charades) | Charades-STA | 20 | relative integers in [0, 100] | 0.1 |
| [`qvhighlights/`](qvhighlights) | QVHighlights | 60 | integer seconds | 1.0 |

## Results

R1@{0.3, 0.5, 0.7} / mIoU, Table 1 of the paper (Charades-STA and DiDeMo evaluated with 20 frames, QVHighlights with 60).

| Source → Target | R1@0.3 | R1@0.5 | R1@0.7 | mIoU |
|---|:-:|:-:|:-:|:-:|
| Charades-STA → Charades-STA | 80.3 | 69.1 | 49.2 | 59.9 |
| Charades-STA → QVHighlights | 77.5 | 59.8 | 36.9 | 55.2 |
| Charades-STA → DiDeMo | 69.2 | 55.4 | 32.3 | 48.5 |
| QVHighlights → Charades-STA | 73.3 | 50.2 | 28.0 | 49.0 |
| QVHighlights → QVHighlights | 89.5 | 82.2 | 69.1 | 75.3 |
| QVHighlights → DiDeMo | 62.4 | 36.0 | 17.4 | 40.6 |

## Usage

Install the [code](https://github.com/KHU-VLL/EVIDENT), then:

```python
from evident.model import load_evident

model = load_evident("mreraser/EVIDENT", subfolder="charades")  # or "qvhighlights"
```

To evaluate:

```bash
torchrun --nproc_per_node=8 evaluate.py --checkpoint mreraser/EVIDENT --subfolder charades \
    --annotation data/annotations/qvhighlights_test.json --video_dir $QVH_VIDEOS \
    --num_frames 60 --output results/cha2qvh.json
```

The checkpoints are released under Apache 2.0. Please also follow the licenses of the datasets they were trained on (Charades-STA is for non-commercial use).

## Citation

```bibtex
@inproceedings{ahn2026evident,
  title     = {EVIDENT: Routing MLLM Adaptation through Entity-Grounded Visual Evidence for Cross-Domain Video Temporal Grounding},
  author    = {Ahn, Geo and Han, Jiwook and Kim, Youngrae and Lee, Joonseok and Choi, Jinwoo},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}
```