multimodalart HF Staff Xunzhuo commited on
Commit
e452ca5
·
1 Parent(s): 960c708

news: add Decision 2.0 (trained) (#38)

Browse files

- news: add Decision 2.0 (trained) (18f467dd687b3f2f97d9277f81e56129b2726208)


Co-authored-by: Xunzhuo <Xunzhuo@users.noreply.huggingface.co>

Files changed (1) hide show
  1. news.html +4 -0
news.html CHANGED
@@ -439,6 +439,10 @@ const ITEMS = [
439
  m: { stars: 0 }, date: '2026-09-22', tags: 'jobe frozen qwen3.5 4b letter logits semif readout no training rtx 3080 local decision index 0.2' },
440
 
441
  /* ---------------- trained ---------------- */
 
 
 
 
442
  { cat: 'trained', title: 'blink: 4B, 9B and 27B typed-decision models', by: { name: 'Govind', handle: null },
443
  desc: 'Three typed-decision LoRA models: Qwen3.5-4B, MiMo-V2.6-Distill-Qwen-9B and Qwen3.8-27B. State plus choice/noul/score questions go in; one forward pass per batch returns option probabilities, no text. Self-hosted TypeSafe wire server. Space serves 4B/9B, not 27B. Code v1.2 Apache-2.0; weights non-commercial.',
444
  links: { gh: 'https://github.com/thegovind/blink', hf: 'https://huggingface.co/thegovind/blink-4b', hf2: 'https://huggingface.co/thegovind/blink-27b', space: 'https://huggingface.co/spaces/thegovind/blink', site: 'https://thegovind.github.io/blink/' },
 
439
  m: { stars: 0 }, date: '2026-09-22', tags: 'jobe frozen qwen3.5 4b letter logits semif readout no training rtx 3080 local decision index 0.2' },
440
 
441
  /* ---------------- trained ---------------- */
442
+ { cat: 'trained', title: 'Decision 2.0: six open decision models, 0.6B to 27B', by: { name: 'vLLM Semantic Router', handle: null },
443
+ desc: 'The second generation of the vLLM Semantic Router decision models, all Apache-2.0: Kai 0.6B (a Qwen3-0.6B-Base fine-tune), Eos 0.8B, Sol 2B and Lux 9B (fine-tunes of Decision 1.0), Nox 4B (merged LoRA on Qwen3.5-4B-Base) and Vega 27B (a LoRA adapter on Qwen3.8-27B). Each answers Choice, Noul and Score questions about one input in a single forward pass, with a probability for every option, through Transformers remote code. Complete self-scored Decision Index 0.2.1 runs are submitted, with the results dataset on the Hub.',
444
+ links: { hf: 'https://huggingface.co/collections/vllm-sr/decision-20-6ab7cf7bdfb506bf8269cb00', hf2: 'https://huggingface.co/vllm-sr/Decision-2.0-Vega-27B', gh: 'https://github.com/apolinario/decision-index/pull/48', site: 'https://huggingface.co/datasets/vllm-sr/decision-2.0-decision-index' },
445
+ m: { hfLikes: 1 }, date: '2026-10-03', tags: 'vllm semantic router decision 2.0 kai eos sol nox lux vega 0.6b 0.8b 2b 4b 9b 27b qwen3 qwen3.5 qwen3.8 fine-tune lora adapter choice noul score one pass system one transformers remote code apache decision index 0.2.1 submission results dataset amd rocm' },
446
  { cat: 'trained', title: 'blink: 4B, 9B and 27B typed-decision models', by: { name: 'Govind', handle: null },
447
  desc: 'Three typed-decision LoRA models: Qwen3.5-4B, MiMo-V2.6-Distill-Qwen-9B and Qwen3.8-27B. State plus choice/noul/score questions go in; one forward pass per batch returns option probabilities, no text. Self-hosted TypeSafe wire server. Space serves 4B/9B, not 27B. Code v1.2 Apache-2.0; weights non-commercial.',
448
  links: { gh: 'https://github.com/thegovind/blink', hf: 'https://huggingface.co/thegovind/blink-4b', hf2: 'https://huggingface.co/thegovind/blink-27b', space: 'https://huggingface.co/spaces/thegovind/blink', site: 'https://thegovind.github.io/blink/' },