hexproj-collab/hexgen-share

hexgen — abc_v1 training data

Training pairs for a decoder that reads a CAD solid and emits a program which compiles to a hexahedral finite-element mesh. Exported from the hexgen-corpora Modal Volume by modal_app/share_corpus_fn.py.

path	contents	size
hexgen_abc_v1/pairs.jsonl	133,238 (part, program) training records, one JSON object per line	2,335 MB
hexgen_abc_v1/holdout_graphs.tar.gz	10,179 .brepgraph.json — held-out CAD parts, never trained on	188 MB raw
hexgen_abc_v1/holdout_tess.tar.gz	10,179 .tess.json — the same parts triangulated, matched by filename	186 MB raw
hexgen_abc_v1/export_summary.json	export provenance	838 B

A .brepgraph.json describes one CAD solid: faces with surface type and parameters, edges with curves, vertices with positions, and the loops tying them together. A .tess.json is that solid as triangles; the mesh-vs-CAD tools take the two as a pair.

uvx hf download hexproj-collab/hexgen-share --repo-type dataset --local-dir ./hexgen-data
cd hexgen-data/hexgen_abc_v1
mkdir -p holdout_graphs holdout_tess
tar xzf holdout_graphs.tar.gz -C holdout_graphs
tar xzf holdout_tess.tar.gz   -C holdout_tess

Do not regenerate the .tess.json files. They were produced by the CAD kernel on x86 and are frozen; tessellation is not reproducible across architectures, so rebuilding them on ARM silently changes every downstream score.

---

hexplan_decision_v1/ + hexplan_p413_28m/ — the P4-13 28M scaling arm

Added 2026-08-15 so the 28M arm of the hexplan scaling probe can run on a Spark other than the one it was queued on.

path	contents	size
hexplan_decision_v1/records.jsonl	the training corpus the arm trains on — 121,648 records	5,990 MB
hexplan_decision_v1/plan_spec.json	the vocabulary/spec the records were encoded against	10 KB
hexplan_p413_28m/p413_28m_inputs.tar.gz	everything else the arm needs: the 200 validation parts' .brepgraph.json and .tess.json, their teacher objects, the archived baseline emission the control leg rescores, and the two id lists	1.2 MB

uvx hf download hexproj-collab/hexgen-share --repo-type dataset --local-dir ./hexgen-data
cd hexgen-data/hexplan_p413_28m && tar xzf p413_28m_inputs.tar.gz
cat p413_28m_inputs/README.md      # how to point the job at these roots

Two things that decide whether the run is readable:

• Run the job at git SHA f92284c3. The control leg rescores the archived baseline emission and must print 114/200 there. At a later SHA it prints 115 — a legitimate experiment, but a different one, and the two must not be pooled.
• Keep the .tess.json sidecars beside the graphs. Without them the mesher builds a different, much worse mesh, which reads as a finding and is an artefact. As above: never regenerate them.