# PAct official-equivalent finetune route 128

- Dataset: 126 GAPartNet instances across 8 categories.
- Preprocess: 660 SS part latents, 534 reconstructed SLAT part latents.
- Stage1: official PartBasedSparseStructureFlowModel, 160 steps, final loss 1.0833; 115 instances after official filtering.
- Stage2: official SLatFlowModel + articulation regression, 160 steps, final loss 7.5800, final art L2 0.0603; 126 instances after sanity check.
- Inference: 8 category-covering samples, 19 exported part GLBs, 8 portal-combined static GLBs, 8 object.json files, 2 grid videos.

Conclusion: official-equivalent finetuning is technically viable and stabilizes the articulation regression signal, but the current generated geometry still shows part collapse/empty-part fallbacks on several categories. This is a real PAct limitation on non-PM GAPartNet-style data, not just a portal issue.