README: version note — weights updated to stage-5 multi-task distill (2026-05-30)
Browse files
README.md
CHANGED
|
@@ -26,6 +26,12 @@ It is the **speech/text backbone** of [ViBES](https://github.com/Juzezhang/ViBES
|
|
| 26 |
speech-language-behavior model) — a lightweight, low-latency alternative to the GLM-4-Voice-9B base.
|
| 27 |
The motion experts are released separately: [`ViBES-Face`](https://huggingface.co/JuzeZhang/ViBES-Face).
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
## Model
|
| 30 |
|
| 31 |
- **Architecture:** same `ChatGLMForConditionalGeneration` family as the teacher, scaled down —
|
|
|
|
| 26 |
speech-language-behavior model) — a lightweight, low-latency alternative to the GLM-4-Voice-9B base.
|
| 27 |
The motion experts are released separately: [`ViBES-Face`](https://huggingface.co/JuzeZhang/ViBES-Face).
|
| 28 |
|
| 29 |
+
> **Version note (2026-05-30):** the weights were updated to our latest distillation checkpoint —
|
| 30 |
+
> the final **stage-5 multi-task** model of the distillation pipeline (covering the S2S / S2T / T2S /
|
| 31 |
+
> T2T modes below). Same architecture, tokenizer, and interface as the initial release; quality is
|
| 32 |
+
> improved. ViBES 0.5B motion experts trained against an earlier snapshot should be retrained or
|
| 33 |
+
> re-validated against this base.
|
| 34 |
+
|
| 35 |
## Model
|
| 36 |
|
| 37 |
- **Architecture:** same `ChatGLMForConditionalGeneration` family as the teacher, scaled down —
|