Skip to content

History / Training the Travel SLM

Revisions

  • docs(slm): travel SLM v2 + v3 results (data scaling + constrained decode) Records the v2 (701-turn data scaling + render tool_summary conditioning) and v3 (950-turn targeted data + logit-level constrained decoding) iterations: tool 0.56→0.80→0.94, widget_type 0.38→0.65→0.79, schema validity held 1.0 (extract now guaranteed by constrained decode). show_places 3/14→27/27. 3B probe aborted (capacity not the bottleneck). Not yet Muno-ready: widget_type 0.79 is the blocker. Adds the batch-1 memory caveat.

    @kadyapam kadyapam committed Jun 27, 2026
  • docs(slm): add Travel SLM journey (track record) + training guide pages Two new wiki pages for the travel-domain SLM: - Travel-SLM-Journey: chronology, the teacher-ceiling finding (constrained decoding beats scale), the v1_constrained dataset, the first real Apple-Silicon MLX LoRA, the honest not-yet-Muno-ready verdict, registry lineage, and the next-iteration plan. - Training-the-Travel-SLM: reproducible engineer guide — dataset build, local MLX + GKE GPU (PEFT) training paths, eval vs the oracle floor, G3 registry lineage, exact reproduce commands. Cross-linked from Home, _Sidebar, and the existing travel-slm page; the travel-slm page gets a Phase A+B status banner. Sourced from ops#216/217/ 219/220, travel#73/#75, the travel-mlx-v1 model card + eval report, and the v1_constrained dataset manifest. Docs only.

    @kadyapam kadyapam committed Jun 27, 2026