Datasets:

---

ParityLab

/

Hand-Datasets-Raw

Like 0

Follow

ParityLab 9

Candidate PR: replace ARCTIC shared-v1.0 with the verified shared-v1.1
release. Please review before merging.

Full local migration: 2973 sequences / 3051 views, validated and audited locally. The PR
adds ARCTIC/shared-v1.1 (55,417 files) and removes ARCTIC/shared-v1.0 (91,713 files) in
one separate commit -- the same replacement every other dataset's v1.1 PR makes: dex-ycb is
already shared-v1.1-only on main, and PR #4 does the same for HOT3D, OakInk-v1,
OakInk-v2 and FreiHAND. ARCTIC/raw is untouched, no other dataset's files or LFS
tracking lines are modified (both re-checked on the candidate branch before merging), and
main keeps shared-v1.0 fully available until this PR merges.

Pushed under yhl's authorization; merging is left to the reviewer.

Why the removal is in this PR, and in this order: the release's own
doc/PUBLICATION_PLAN.md (section 本树当前保留了一份 shared-v1.0) registers the duplicate
tree as a deliberate, temporary deviation from the registry's release_layout and
prescribes exactly what this PR does -- the upload adds shared-v1.1 only, and the old
tree's removal is a separate change made after the publication manifest and every path,
size and content digest have verified. That file is hashed inside the release, so it could
not be edited without re-validating the 737 GiB tree; this PR follows it rather than
rewriting it. The order is enforced by the tool that pushed this branch: the 55,417 new
paths were all verified against the plan before the removal commit was created.

Conventions this release carries deliberately, called out here because
doc/PUBLICATION_PLAN.md is a hashed file inside the release and could not be edited
without re-validating the 737 GiB tree:

• doc/validation/ holds 54 sampled per-view review pages (6 sequences x 9 cameras, seed
  20261014. plus index.html and V1_1_VALIDATION.json. They are acceptance evidence,
            not training input. The published H2O and dex-ycb releases ship the same kind of pages.
• The release tree ships no upload script, per doc/PUBLICATION_PLAN.md; the tool used
  for this PR lives outside the release, in the dataset workspace's tmp/arctic-tools/.
• .gitattributes: this repo tracks LFS with explicit per-path lines, and at main that
  file is 10,484,654 B against the Hub's 10 MiB ceiling for non-LFS files -- 1,106 B of
  headroom. ARCTIC's own 5,535 lines are exactly its .mkv (3,051) and .obj (2,484)
  paths, the two types no global pattern covers, so the release cannot be added path by
  path: the Hub appends a line for every new LFS path (the candidate commit's two proof
  copies already added 9 lines / 992 B). This PR uses the glob form the repository already
  carries for dex-ycb/shared-v1.1/**/*.obj (live on main) and for HOT3D / OakInk-v1 /
  OakInk-v2 / FreiHAND (PR #4), leaving ~634 KB of headroom. With the old tree removed the
  file keeps only the two ARCTIC/shared-v1.1/** globs -- deliberately not the H2O /
  dex-ycb shape of stripping the dataset's lines outright, because the Hub appended no
  per-path lines for this release (they went up while the globs were live) and those two
  lines are what keeps its .mkv and .obj files LFS-tracked for a fresh clone. Every
  non-ARCTIC line is byte-identical, re-checked on the candidate branch before merging.
• File counts, since the point of this migration is fewer, larger files: ARCTIC goes from
  91,713 files for 3,051 views (30 per view) to 55,417 (18 per view). The whole reduction
  is the per-frame numeric tensors merged into one .pt per view -- .pt 57,065 -> 20,635,
  with every other type unchanged. The repository ends up 36,296 files smaller than before
  this PR, not 55,417 larger: the old tree is not left behind.

The release's own status flags (human_full_review: pending, publication_authorized: false, hf_writes: false) are untouched: they describe human review state, which is
recorded outside the tree.

🤖 Generated with Claude Code

复核意见的修改已经推上本候选分支（未合并）。

修的是你点出的那一类洞：一次「选中的集合为空」的运行仍可能报通过。

• script/arctic_v1_1/validate.py：选中的视图与发布集不相交时不再对空集报 pass，按错误处理；子集运行不得写发布级验收报告（doc/validation/validate_report.json），也不得重建清单——不冒充完整验收；--expect-views 的语义随之收口到「本次选择自己的范围」。
• script/arctic_v1_1/verify.py：同一条规则落到逐视图校验（源里没有的序列、尚未迁移的序列各有明确处置）。
• doc/validation/validate_report.json 由修好的校验器重新产出（报告本身带着 recipe.code_sha256，描述的就是本树这些工具的字节），doc/FULL_MIGRATION_SUMMARY.json 与 script/arctic_v1_1/README.md 据此更新。

推法是最小重推，这一点可以逐条核：本地树与分支自己的文件索引（HF_PUBLICATION_PLAN.json 记着每个路径的 size 与内容身份）逐路径对比，只提交字节真正不同的路径——5 个，其余 55,412 个路径一个字节都没有重传，媒体（737 GiB）与 doc/validation/ 下 54 个逐视图复核页未动。两条守卫使它可信：计划描述的那个 head 必须仍是分支的 head（不是就停下）；本地树必须仍是审计过的那棵树（与树外清单逐条比）。

回归测试 tmp/arctic-tools/test_v1_1_review_fixes.py 15 条全过（含一次走默认路径的完整验收调用——那次要把整棵树哈希一遍）。修正之后本地树重新跑过校验与清单重建，所以分支上的字节就是审计过的字节。推送后逐路径从 Hub 重读复核，另起进程的 tmp/arctic-tools/verify_pr5_state.py 也重跑通过（PR5_STATE_OK）。

仍停在合并之前：human_full_review: pending、publication_authorized: false、hf_writes: false 一律未动。

另有一件需要你定的事：Hub 提示本 PR 与 main 有一个 .gitattributes 冲突。main 在我们复核期间被别的数据集推着前移（本 PR 的 base 是 54c1bc81，base 72,046 行 / 10,484,654 B，main 现 68,934 行 / 9,822,963 B）。

这不是本 release 的缺陷：发布器验的是「相对本 PR 的 base commit 只动了 ARCTIC 自己那几行」，成立，且分支上其他数据集的 LFS 行与 base 逐字节相同。冲突是「谁最后合并谁负责把两边合起来」这件事本身。

合并期的修法已经算好、没有推送（在 tmp/arctic-tools/RECONCILED_GITATTRIBUTES，做法是以当前 main 为底套用本 release 自己那两处改动——ARCTIC 的逐路径行换成 shared-v1.1 的两条 glob、v1.0 的 glob 随旧树去掉）：9,189,734 B / 63,401 行，其他数据集的行与当前 main 逐字节相同（脚本会断言），提交后离 10 MiB 上限还有 1,296,026 B 余量（main 现在只剩 662,797 B，所以合并反而更宽松）。要不要这么做、什么时候做，交你定；细节在 tmp/arctic-tools/PR5_MERGE_NOTES.md。

Review fixes applied and tested; current main LFS rules reconciled without changing other datasets or raw data. Regression checks pass and content manifests refreshed. Data identity is preserved; historical semantic exceptions and pending human review remain explicitly documented in doc/REVIEW_FIXES.md. This merge does not constitute human visual acceptance.