athrael-soju/HydraQwen3.5-4B
Visual Document Retrieval • Updated • 10
Dual-head VLM: ColBERT retrieval + autoregressive generation by toggling one LoRA. Canonical 4B + 0.8B, omni proof-of-concept, baselines.
Note Canonical Hydra-4B (paper version): 4.60B params, dim=320, multilingual.
Note Hydra-0.8B (small-scale instantiation; has Gradio demo Space).
Note Hydra-Omni: proof-of-concept on Qwen2.5-Omni-3B for image/audio/video retrieval + speech generation.
Note GritLM-style joint-training ablation (paper Sec. 5).