Mizar — Audio-Language Model Family - a KaiyangLi Collection

KaiyangLi 's Collections

updated Sep 23

Compact ALMs at sub-200M and 3B scales, leading among evaluated size-matched baselines on MMAU, MMAR and ADQA-dev (clean split).

KaiyangLi/Mizar-159M

Audio-Text-to-Text • Updated about 1 month ago • 3

Note 159.3M parameters; CED-Small + SmolLM2-135M; three-stage training. License: BSD-3-Clause-Clear. Code: https://github.com/KaiyangLi1992/Mizar_159M
KaiyangLi/Mizar-3B

Updated about 1 month ago • 15 • 2

Note Ke-Omni-R-3B with RT-OPD; released LoRA adapter, seed 85 / step 626. License: BSD-3-Clause-Clear; base-model terms apply. Code: https://github.com/KaiyangLi1992/RT-OPD