lipeize/collsplit-v15-stability

collsplit_v15 stability label — PRIVATE TRANSFER ARTIFACT

Do not redistribute. This repository is a private machine-to-machine transfer artifact, not a dataset release. Its contents are derived from GraspNet-1Billion, which is licensed for research use; redistributing this data publicly, or to anyone not already covered by that license, is not permitted. The repository is private and must stay private.

Companion repos: lipeize/collsplit-v15-sidecar carries the collision_label/ and meta/ sidecar whose row space this label lives in; lipeize/collsplit-v15-width carries the width campaign's output, row-aligned with this one. This repo carries the stability campaign's output — the third and final label class of collsplit_v15.

What is here

File	Contents
stability_label_XXXX.tar.zst	one scene dir XXXX/ of stability_label/, holding stability_YYYY.npz (one per view)
manifest.json	per shard: scene id, file count, tar byte size, sha256 of the tar
qa.tar.zst	the qa/ directory — the campaign's QA sweep reports and acceptance verdicts

190 scene shards / 48,640 views / 362.4 GiB. The source tree is 362.6 GiB apparent (362.6 GiB on disk); the .npz payloads are already deflate-compressed, so zstd -1 only removes tar padding — budget 362.4 GiB for a full pull. manifest.json is the byte authority.

Every scene holds exactly 256 views — including (41, 255). The width repo excludes that one view (v1 pose-gate inheritance); the stability resampler reads each view's own annotation poses directly, so the gate has no analogue here and nothing is excluded. Do not carry width's (41, 255) special case into this label; the aligned-view intersection of the two corpora is width's 48,639.

Format contract

Read the authoritative spec before consuming these labels: docs/panel_impl/stability_v15.md on main in the graspformer repo. Summary:

stability_label/<scene:04d>/stability_<view:04d>.npz

key	shape / dtype	semantics
p	(N,300,12,4) uint8	rint(clip(p,0,1)*255) of p = 1.1 - mu; 0 = invalid (mu = -1) or all-invalid stamp
meta	json	version collsplit_stability_v15, mapping contract, encode, alive convention, n_rows, n_stamped_all_invalid

Three things are easy to get wrong:

1. The alive threshold is p >= 0.2 (u8 >= 51), i.e. mu <= 0.9. The v14 habit (threshold 0.1) silently promotes mu = 1.0 cells under this mapping: the weakest real grade encodes to p = 0.1 — deliberately one ordinary grade step above invalid, with no sentinel cliff.
2. The stored label is raw. No gap remap, no tie-breaker blur, no thresholds are baked in — those are train-time transforms, specified (with their ruled hyperparameters) in the spec. Any alive/dead decision reads this raw channel.
3. All-zero rows are real data. Points whose every raw annotation score is -1 skip interpolation and are stamped all-zero: 5,125,595 rows (2.6%) corpus-wide, per-view count in meta (n_stamped_all_invalid). Rows that mix -1 with real grades DO interpolate — their p = 0.1 cells are signal.

Row space N is the collision sidecar's: all un-occluded points of the view, in gid order — row-for-row aligned with width_label by construction, with no viability filtering (v14's score_thresh row-dropping does not run). Corpus total 198,290,436 rows over 48,640 views. The trailing (300, 12, 4) axes are the project's standard 4D grasp cell space: 300 view directions x 12 in-plane rotations x 4 gripper depths.

QA acceptance

QA ran in three waves behind the campaign (scenes 0-100 / 101-160 / 161-189); sweep reports and acceptance verdicts ship in qa.tar.zst. Rulings: wave 1 corpus ACCEPTED (stab_qa_20260810_bar4_analysis.json), wave 2 scenes 101-160 ACCEPTED (stab_qa_wave2_verdict.json), wave 3 scenes 161-189 ACCEPTED — v15 stability COMPLETE 190/190 (stab_qa_wave3_verdict.json).

check	result	file
format cert (bar 1)	PASS — 0 fails over all 48,640 views (25,856 + 15,360 + 7,424)	stab_qa_*.json, wave verdicts
all-invalid stamp (bar 3)	PASS — exhaustive stamp-set equality vs the raw census, 0 fails, all waves	stab_qa_*.json, wave verdicts
brute probe (bar 4)	accepted one-sided — the pre-registered bar failed on a probe mis-calibration; the final mechanism-derived bar "overshoot <= one grade quantum" has zero violations across all waves (see below)	stab_qa_20260810_bar4_analysis.json, wave verdicts
histogram (bar 5)	PASS after recalibration — wave 1: one boundary outlier; wave 2: 344 outliers attributed to bound mis-calibration (richness-scaled interpolation spread, uniform corpus-wide); wave 3 under the delta-calibrated bound: 0/7,424	wave verdicts
determinism (bar 6)	PASS — 24/24 sampled scenes bit-identical on regeneration (12 + 6 + 6 across waves, both compute-path and fastpath populations)	determinism_*.json

Bar 4 as pre-registered failed, and the raw sweep files say so — read this before ranking by peak stability. The registered probe (raw-peak preservation, ok_frac >= 0.99) measured 0.605 in wave 1. Attribution: the proxy assumed RBF transport preserves sharp per-row peaks, but the interpolant attenuates them — deviations are pure undershoot (overshoot beyond 0.06 at 0.1% of rows, p95 +0.000). The accepted replacement bar is one-sided (stored_row_max <= raw_best + 0.06; final form "overshoot <= one grade quantum", zero violations across all waves — wave-3 verdict). Row-max attenuation is a real, recorded property of these labels (median -0.043, p5 -0.349 on sharp-peak rows): a consumer ranking points by absolute stored peak height inherits it. The sweep jsons carry pass_sweep: false under the pre-registered bars; the *_analysis.json / *_verdict.json files are the acceptance rulings.

Download and reconstruct

REPO=lipeize/collsplit-v15-stability
export HF_HUB_ENABLE_HF_TRANSFER=1

huggingface-cli download "$REPO" --repo-type dataset \
  --local-dir /datadir2/hf_incoming_stability \
  --include 'manifest.json' 'qa.tar.zst' 'stability_label_*.tar.zst'

Subsets pick up scenes by glob, e.g. 'stability_label_00[0-4][0-9].tar.zst' for scenes 0000-0049. manifest.json and qa.tar.zst are ~100 KiB together and worth taking every time.

Verify against the manifest BEFORE untarring — a truncated shard untars into a plausible-looking partial scene, which is far more expensive to notice later than a hash check now:

cd /datadir2/hf_incoming_stability
python - <<'PY'
import hashlib, json, pathlib, sys
man = json.load(open("manifest.json"))
bad, seen = [], 0
for name, rec in man["shards"].items():
    p = pathlib.Path(rec["file"])
    if not p.exists():
        continue                       # shard you chose not to download
    seen += 1
    if p.stat().st_size != rec["tar_bytes"]:
        bad.append(f"{p}: size {p.stat().st_size} != {rec['tar_bytes']}"); continue
    h = hashlib.sha256()
    with open(p, "rb") as f:
        for c in iter(lambda: f.read(8 << 20), b""):
            h.update(c)
    if h.hexdigest() != rec["sha256"]:
        bad.append(f"{p}: sha256 mismatch")
print("\n".join(bad) if bad else f"all {seen} present shards verified")
sys.exit(1 if bad else 0)
PY

Only then unpack:

DEST=$GRASPNET_SAVE_DIR/collsplit_v15
mkdir -p "$DEST/stability_label"
for t in stability_label_*.tar.zst; do
  nice -n 10 tar -I zstd -xf "$t" -C "$DEST/stability_label"
done
tar -I zstd -xf qa.tar.zst -C "$DEST/stability_label"   # restores stability_label/qa/

Member paths are bare XXXX/. Every shard in this repo — like every shard in lipeize/collsplit-v15-sidecar and lipeize/collsplit-v15-width — stores members as XXXX/<file> with no class prefix. Untar this repo's shards into a dedicated stability_label/ destination, never into the same directory as the sidecar's collision_label_*/meta_* shards or width's width_label_* shards: the scene dirs would merge and the label sets would become indistinguishable in place.

Sanity check after unpacking — every scene holds 256 views, no exceptions:

find "$DEST/stability_label" -mindepth 1 -maxdepth 1 -type d -not -name qa \
  | while read d; do echo "$(basename $d) $(ls $d | wc -l)"; done | awk '$2!=256'
# expect NO output

Integrity

manifest.json carries a sha256 per shard, computed on the tar in the same pass that wrote it, before upload. After upload, every shard's remote byte size AND remote LFS sha256 were verified against that manifest entry; three scene shards were additionally re-downloaded whole and byte-compared against their local staging tars, and scene 0041 was round-trip untarred and byte-compared against the read-only source. Verify after download and before untarring, as above.