Alexander Korolev c2871f857a kindling 4x re-bench on 27cf839 image (label kindling-4x-27cf839): 94/100 (166/176), identical to the a3c3a1d baseline — same fails TC-43/51/68/80, same depth curve. The megamoe #67 + spec-config #64 in the new image neither regress nor improve GLM-5.3-Flash 4x quality. Model now floats via refs/main (da920bb0 at bench time); manifest predates the models-field, so it carries no snapshot (next bench will). před 3 týdny
..
reports c2871f857a kindling 4x re-bench on 27cf839 image (label kindling-4x-27cf839): 94/100 (166/176), identical to the a3c3a1d baseline — same fails TC-43/51/68/80, same depth curve. The megamoe #67 + spec-config #64 in the new image neither regress nor improve GLM-5.3-Flash 4x quality. Model now floats via refs/main (da920bb0 at bench time); manifest predates the models-field, so it carries no snapshot (next bench will). před 3 týdny
recipe.toml 24d6737bc3 de-pin model revisions: kindling 2x/4x model -> floating org/model id (engine resolves via NAS refs/main -> latest snapshot, same as every other recipe; the pinned da920bb0 stays as an informational comment); 6x notes its tp6pad source rev. Bench manifest now records the model's resolved refs/main snapshot so a floating model stays reproducible (backward-compatible #[serde(default)]). před 3 týdny