news: add Vega 0.8B and 4B (vega-ml), from the creator of Laya
Adds one trained entry for Vega (brand: vega-ml), from the creator of Laya, and lists its Hub repo in the Space card.
What it is. One typed-decision engine in two sizes, 0.8B and 4B, as adapters over frozen Qwen3.5 backbones (public_intents and typed_decisions). Single pass per question, a probability on every supplied option, an answer on every row. Apache-2.0. Both sizes are in one Hub repo, the 4B under 4b/.
Where the numbers come from. One run on 2x T4 against the live hosted Jev endpoint over 43 public tasks, JevBench and DecisionBench, 3,841 decisions per system. These are this repo's own measurements against the endpoint on those rows, not Decision Index results:
| Vega 0.8B | Vega 4B | Jev | |
|---|---|---|---|
| public macro accuracy | 0.556 | 0.717 | 0.836 |
| public NLL | 1.099 | 0.768 | 0.878 |
| DecisionBench ECE | 0.103 | 0.098 | 0.122 |
| DecisionBench soft-gold KL | 1.012 | 0.854 | 6.099 |
| DecisionBench coverage | 1.000 | 1.000 | 0.997 |
| median latency per decision | 74 ms | 106 ms | 162 ms |
The notebook that produced them is public, pins its library versions, derives its tables from the run JSON rather than typed-in numbers, and reruns in about 4.5 hours on 2x T4. The long-form write-up includes an honest accounting of how thin most of the per-task wins are (eight of twelve are three rows or fewer).
Links in the entry: Hub repo, GitHub, write-up.
Happy to correct anything, including the category, if trained is not the right bucket for adapters on a frozen backbone. A Decision Index submission for both sizes through the reproduction kit is next, separately.
Also fixes nothing else: news.html gains one entry after the Laya one, and README.md gains one line in models:.