Vela Modality
Vela Modality identifies whether a request needs text output (AR), image generation (DIFFUSION), or both (BOTH).
307M parameters · Input capacity: 32,768 tokens, including special tokens.
Evaluation
Macro F1 (×100) on the same development sets, compared with the original mmBERT32K Modality model. Higher is better.
| Evaluation | Original mmBERT | Vela |
|---|---|---|
| Short inputs · 929 | 38.40 | 81.10 |
| Controlled 4K context · 18 | 26.67 | 82.22 |
| Controlled 8K context · 18 | 26.67 | 82.22 |
| Controlled 16K context · 18 | 26.67 | 82.22 |
| Controlled 32K context · 18 | 26.67 | 82.22 |
Short inputs cover six languages. Long-context evaluation uses six English/Chinese examples placed at three positions and repeated across lengths. Both models process complete inputs in FP32. These development sets informed Vela selection; they are not an independent test set.
Quick start
With PyTorch and Transformers 4.57.6 or 5.17.0:
from transformers import pipeline
model_id = "vllm-sr/Vela-1.0-Encoder-307M-Modality"
model = pipeline("text-classification", model=model_id, device=-1)
print(model("Create a watercolor image of a lighthouse at sunset.", top_k=None, truncation=False))