Brilliant work on OneVision-Encoder! 🎉
I'm trying to reproduce the LLaVA-NeXT-Video evaluation results following the instructions in the README.
For video benchmarks (e.g., VideoMME), I ran:
TASKS="videomme" bash scripts/eval/eval_ov_encoder.sh
However, I noticed this line in the script:
MODEL_PATH="${MODEL_PATH:-trained_model/must_contain_llava_in_name}"
I've searched through the repository and the Hugging Face organization (lmms-lab-encoder / lmms-lab), but I couldn't find a released model checkpoint whose name contains "llava".
❓ Could you clarify:
Am I misunderstanding the evaluation workflow?
Or is the LLaVA-integrated checkpoint not yet publicly released?
Any guidance would be greatly appreciated! Thanks again for the amazing work. 🙏
Brilliant work on OneVision-Encoder! 🎉
I'm trying to reproduce the LLaVA-NeXT-Video evaluation results following the instructions in the README.
For video benchmarks (e.g., VideoMME), I ran:
TASKS="videomme" bash scripts/eval/eval_ov_encoder.sh
However, I noticed this line in the script:
MODEL_PATH="${MODEL_PATH:-trained_model/must_contain_llava_in_name}"
I've searched through the repository and the Hugging Face organization (lmms-lab-encoder / lmms-lab), but I couldn't find a released model checkpoint whose name contains "llava".
❓ Could you clarify:
Am I misunderstanding the evaluation workflow?
Or is the LLaVA-integrated checkpoint not yet publicly released?
Any guidance would be greatly appreciated! Thanks again for the amazing work. 🙏