Instructions to use gavinlaw/multimodal-understanding-ste-adapters with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use gavinlaw/multimodal-understanding-ste-adapters with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Experimental STE pilot adapters
Eight rank-8 language q/k/v/o LoRA adapters, four arms and seeds 71201/71202. Load the matching immutable Qwen3-VL-4B base with PEFT. Evidence and joint arms add head.pt, loaded with EvidenceHead from soft_temporal_evidence.py. Head files in answer/sequence arms are unused initializations, not learned evidence models. Training and reproduction code, complete development predictions, and limitations: https://huggingface.co/datasets/gavinlaw/multimodal-understanding-ste-outputs CaptainCook4D original dataset: https://captaincook4d.github.io/captain-cook/ These are development experiments, not generally validated models. No base model weights or source video bytes are included. Retain upstream Apache-2.0 notices.
- Downloads last month
- -
Model tree for gavinlaw/multimodal-understanding-ste-adapters
Base model
Qwen/Qwen3-VL-4B-Instruct