Commit History

  1. Results tab: three corpora — add held-out OlympiadBench (58.1%) alongside PRM800K (~0.1%) and SciBench (83.4%)

    tomyimkc
    a36d178

  1. Results tab: lead with the measured boundary (~0.1% NL reasoning vs 83.4% unit-bearing), demote the planted 67/67 to a labelled self-test

    tomyimkc
    3a2ec86
  2. Wire Self-Judge + Step-Verify; stop the step checker silently passing wrong arithmetic⁠

    tomyimkc
    8f5fb99

  1. Speed: 2 models (drop slow DeepSeek), max_tokens=400, timeout=30s

    tomyimkc
    1db7af5
  2. Redesign: multi-model parallel UI — one prompt, 3 LLMs, side-by-side comparison

    tomyimkc
    83a00a9
  3. Fix all 6 issues: LaTeX strip, coefficients regex, circular multi-word, Faraday unit, step checker LaTeX, bare number extraction

    tomyimkc
    e61c915
  4. Fix: auto-fallback models (Qwen→Llama-3.3-70B→DeepSeek-V3)

    tomyimkc
    46e62f2
  5. Fix: strip LaTeX markup before extraction (\boxed, \text, etc.)

    tomyimkc
    635f12c
  6. Fix: step_checker handles dict steps from _parse_cot_steps

    tomyimkc
    f0e80d1
  7. Initial deploy: polished demo with Qwen2.5-7B API + step verification + 7 challenges

    tomyimkc
    b6e520e
  8. initial commit

    Verified signature
    e2f52d5