Commit Graph

  • 8b567cb531 chat mode: json_output toggle to return clean extracted JSON main Ethanfel 2026-07-02 02:09:36 +02:00
  • f7ea559690 Speed: auto flash-attention/SDPA + document perf levers Ethanfel 2026-06-27 11:18:11 +02:00
  • 22fd24b29e Re-enable reasoning for accurate verdicts (no-think rubber-stamped 'match') Ethanfel 2026-06-27 10:56:47 +02:00
  • fee136e98c Expose long-text fields as input sockets (forceInput) Ethanfel 2026-06-27 10:49:37 +02:00
  • 0e9e99b8b2 Handle reasoning models (Qwen3.5/3.6): no-think + JSON-only + prose fallback Ethanfel 2026-06-27 10:25:16 +02:00
  • f5be04a5cb Fix: render text inputs as editable fields, not input sockets Ethanfel 2026-06-27 10:18:12 +02:00
  • d389d6daff Trim dead inputs: drop fp16 precision and prompt_used Ethanfel 2026-06-27 10:03:06 +02:00
  • 271aa8ae42 Add chat mode: use the node as a general VLM, not just a judge Ethanfel 2026-06-27 09:55:36 +02:00
  • 5cff883914 Add Qwen3.5/3.6 abliterated (safetensors) + split model/quant selectors Ethanfel 2026-06-27 09:50:13 +02:00
  • e29df0b319 Keep node transformers-only: drop GGUF presets from dropdown Ethanfel 2026-06-27 09:26:32 +02:00
  • 34adb946a4 Add model dropdown (presets w/ VRAM) + manual override; list HauhauCS GGUF models Ethanfel 2026-06-27 09:24:28 +02:00
  • 887dfc0bbb Add analysis profiles with distance/proximity-aware axes Ethanfel 2026-06-27 00:48:46 +02:00
  • 06992506d7 Drop named-position axis for grounded geometry (30B still mis-names positions) Ethanfel 2026-06-26 23:49:23 +02:00
  • e4dfaac63b Correct 4B 'partial' bias on identical values; harden verdict rule; note model-capability limits Ethanfel 2026-06-26 23:43:34 +02:00
  • 69c1d6deb4 describe emits one canonical reference; compare can anchor on it Ethanfel 2026-06-26 23:22:57 +02:00
  • 53f1f9b9b4 Switch compare to discrete verdicts + granular pose axes + per-axis definitions Ethanfel 2026-06-26 23:15:51 +02:00
  • c7ef756a71 Add describe (first-pass) mode to the judge node Ethanfel 2026-06-26 23:04:09 +02:00
  • 959ec70065 Redesign judge output for calibration: per-axis {score, ref, gen}, drop local fix suggestions Ethanfel 2026-06-26 22:52:40 +02:00
  • aa3983d94a Fix: handle missing processor chat template Ethanfel 2026-06-26 22:36:39 +02:00
  • 95198a15b5 Initial commit: VLM-as-judge prompt calibration loop Ethanfel 2026-06-26 22:15:56 +02:00