Post
84
Jev-style decisions, v4.2: Jev-Style-2B-Decision-v4.2
A text or JSON state, optional images, and typed questions in; a probability for every supplied option out. Choice, yes/no or an ordered score, without parsing a generated reply.
Try the new browser demo:
chaoliangUNSW/jev-style-2b-v4-2
• Decision Index 0.3 public index: 46.74, up from 44.37 for our 2B v4 (+2.37 points on the same public suite).
• One complete self-run with the kit's runner and scorer: 140,178 scoreable requests, all successful. An independent CPU rescore reproduced 46.74.
• Up to 32,768 input tokens, image tokens included, and 255 options. Over-budget inputs are refused rather than truncated.
• Several questions about one state share its text computation. PyTorch reads text and images; GGUF and MLX are text-only.
The public score is from the original BF16 CUDA run. The demo uses FP32 on shared ZeroGPU. Submitted for review:
https://github.com/apolinario/decision-index/pull/389
The board's Full score comes from the maintainers' private tests; a public self-run is not a board rank.
Models and matching runtimes:
chaoliangUNSW/Jev-Style-2B-Decision-v4.2
chaoliangUNSW/Jev-Style-2B-Decision-v4.2-GGUF
chaoliangUNSW/Jev-Style-2B-Decision-v4.2-MLX
Results: chaoliangUNSW/decision-index-results-v4.2
Apache-2.0 weights; training sources, data terms and diagnostic-release details are on the model card. Trained in part on soft labels distilled from TypeSafe Jev 1.14 via its public API. Not affiliated with TypeSafe or Jev.
A text or JSON state, optional images, and typed questions in; a probability for every supplied option out. Choice, yes/no or an ordered score, without parsing a generated reply.
Try the new browser demo:
chaoliangUNSW/jev-style-2b-v4-2
• Decision Index 0.3 public index: 46.74, up from 44.37 for our 2B v4 (+2.37 points on the same public suite).
• One complete self-run with the kit's runner and scorer: 140,178 scoreable requests, all successful. An independent CPU rescore reproduced 46.74.
• Up to 32,768 input tokens, image tokens included, and 255 options. Over-budget inputs are refused rather than truncated.
• Several questions about one state share its text computation. PyTorch reads text and images; GGUF and MLX are text-only.
The public score is from the original BF16 CUDA run. The demo uses FP32 on shared ZeroGPU. Submitted for review:
https://github.com/apolinario/decision-index/pull/389
The board's Full score comes from the maintainers' private tests; a public self-run is not a board rank.
Models and matching runtimes:
chaoliangUNSW/Jev-Style-2B-Decision-v4.2
chaoliangUNSW/Jev-Style-2B-Decision-v4.2-GGUF
chaoliangUNSW/Jev-Style-2B-Decision-v4.2-MLX
Results: chaoliangUNSW/decision-index-results-v4.2
Apache-2.0 weights; training sources, data terms and diagnostic-release details are on the model card. Trained in part on soft labels distilled from TypeSafe Jev 1.14 via its public API. Not affiliated with TypeSafe or Jev.