LvW-bench v0.1.0

Changelog

0.1.0 (2026-11-05)

First release.

• Three tracks: decide (System 1, typed decisions scored on accuracy, KL, Brier and ECE),
  answer (System 2, thinking off) and reason (System 2, thinking on).
• Five original System 1 workflows with rule gold. Permissive subsets of GSM8K, MMLU, MMLU-Pro,
  TruthfulQA and IFEval, plus a seeded generated set of 225 items.
• Model inputs: HF links and ids, local .gguf, local .safetensors folders and adapters, and
  OpenAI-compatible servers and Ollama. GGUF runs on llama.cpp (native /v1/systemone for
  decision models). Safetensors run through a uv/transformers helper.
• lvwb CLI and TUI, a Wails desktop app, and a browser runner (wllama, WebLLM).
• Hugging Face dataset with eval.yaml, a leaderboard and seed results for the Spot Snake and
  Mermaid models plus small baselines.