Release v1.0.0: verified Level 1 preprocessing assets

CHANGELOG.md

@@ -0,0 +1,5 @@
+# Changelog
+
+## v1.0.0 — 2026-09-17
+
+Initial Level 1 release: nine processed-text tables, two static-video tasks, original-name recovery, metadata indexes, and SHA-256 verification. No source text rewriting or video re-encoding.

README.md

@@ -0,0 +1,122 @@
+---
+pretty_name: RoboSteer Preprocessing
+language:
+- en
+tags:
+- robosteer
+- preprocessing
+- motion-generation
+- intermediate-assets
+configs:
+- config_name: level1_txt_gen
+  data_files:
+  - split: data
+    path: level1/text/txt_imit/txt_gen/data.parquet
+- config_name: level1_txt_comp_hands
+  data_files:
+  - split: data
+    path: level1/text/txt_imit/txt_comp_hands/data.parquet
+- config_name: level1_txt_comp_legs
+  data_files:
+  - split: data
+    path: level1/text/txt_imit/txt_comp_legs/data.parquet
+- config_name: level1_txt_fore
+  data_files:
+  - split: data
+    path: level1/text/txt_pred/txt_fore/data.parquet
+- config_name: level1_txt_retro
+  data_files:
+  - split: data
+    path: level1/text/txt_pred/txt_retro/data.parquet
+- config_name: level1_txt_inter
+  data_files:
+  - split: data
+    path: level1/text/txt_pred/txt_inter/data.parquet
+- config_name: level1_img_txt
+  data_files:
+  - split: data
+    path: level1/text/img_txt/data.parquet
+- config_name: level1_mul_bal_txt
+  data_files:
+  - split: data
+    path: level1/text/mul_bal_txt/data.parquet
+- config_name: level1_audio_motion_instructions
+  data_files:
+  - split: data
+    path: level1/audio/motion_instructions/data.parquet
+- config_name: level1_img_txt_human_index
+  data_files:
+  - split: data
+    path: level1/image/img_txt_human/index.parquet
+- config_name: level1_img_txt_skel_index
+  data_files:
+  - split: data
+    path: level1/image/img_txt_skel/index.parquet
+---
+
+# RoboSteer Preprocessing
+
+Reusable intermediate preprocessing assets for RoboSteer. This initial release contains Level 1 processed text instructions and image-conditioned static videos. Other levels and preprocessing stages can be added under their own directories.
+
+## Release v1.0.0
+
+- **213,044 text records** in nine lossless UTF-8 Parquet tables.
+- **28,666 original MP4 files**: 14,333 IMG_TXT_HUMAN and 14,333 IMG_TXT_SKEL.
+- Independent, uncompressed tar shards contain MP4 files and their normalized JSON metadata.
+- Per-sample indexes, release inventory, SHA-256 checksums, and validation report.
+
+These are preprocessing assets, not a declared train/validation/test split. The `data` split means the complete asset collection for a configuration. Audio motion instructions are text derived from audio; audio recordings and raw transcripts are not included. MUL_BAL is included as processed text; this release does not contain MUL_BAL videos. Prediction-video manifests are outside this release scope.
+
+## Layout
+
+```text
+level1/
+  text/txt_imit/{txt_gen,txt_comp_hands,txt_comp_legs}/data.parquet
+  text/txt_pred/{txt_fore,txt_retro,txt_inter}/data.parquet
+  text/{img_txt,mul_bal_txt}/data.parquet
+  audio/motion_instructions/data.parquet
+  image/{img_txt_human,img_txt_skel}/
+    videos-00000-of-NNNNN.tar
+    videos-00000-of-NNNNN.index.parquet
+    index.parquet
+  manifests/{assets,validation,provenance}.json
+scripts/restore_text.py
+scripts/build_dataset.py
+SHA256SUMS
+CHANGELOG.md
+```
+
+## Load text or an index
+
+```python
+from datasets import load_dataset
+ds = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_txt_gen", split="data", revision="v1.0.0")
+video_index = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_img_txt_human_index", split="data", revision="v1.0.0")
+```
+
+Text columns: `sample_id`, `source_sample_id`, `task`, `level`, `source_modality`, `text`, `original_relpath`, `size_bytes`, `sha256`. Original UTF-8 bytes, including line endings, can be reconstructed with `text.encode('utf-8')`. Use `scripts/restore_text.py` to restore the original relative file layout.
+
+## Download video shards
+
+```python
+from huggingface_hub import snapshot_download
+snapshot_download("PhoebeCC/RoboSteer-Preprocessing", repo_type="dataset",
+    revision="v1.0.0", local_dir="robosteer-preprocessing",
+    allow_patterns=["level1/image/img_txt_human/*", "level1/manifests/*", "README.md", "SHA256SUMS"])
+```
+
+Each tar is independently extractable. The index identifies its repository-relative `shard`, MP4 `member_path`, JSON `metadata_member_path`, and checksums. Video indexes are searchable tables; this release does not configure an embedded-video viewer. Extract each task into a separate folder. MP4 files are preserved byte for byte, without re-encoding.
+
+## Provenance and alignment
+
+`sample_id` preserves the source filename stem. `source_sample_id` removes only the known task prefix (or audio suffix). Use `(task, sample_id)` as an asset key; use `source_sample_id` to join related tasks, without assuming all tasks cover the same samples. Coverage results are recorded in the validation report.
+
+Video metadata originates from existing sidecar JSON files. Metadata text is the original sidecar description, not necessarily the rewritten `text/img_txt` instruction. Image paths are converted to paths relative to the external public-data root. Source duration and generated duration are metadata values, not newly measured properties. The source generator specifies 25 FPS; this packaging run does not probe actual MP4 frame rates. Source images, external task JSONs, model weights, raw audio, audit logs, and organization backups are not bundled.
+
+## Rights and citation
+
+The repository owner has not yet supplied a release license or complete upstream attribution. No blanket open-source license is asserted for these derived assets. Upstream restrictions remain applicable; consult the owner about permitted redistribution and use. Formal benchmark citation and upstream dataset attribution will be added when supplied.
+
+## Reproducibility
+
+See `level1/manifests/provenance.json` for packaging rules and generator fingerprints, `assets.json` for counts, and `validation.json` for checks. `SHA256SUMS` covers all payload and documentation files except itself. SHA-256 verification can be performed with `sha256sum -c SHA256SUMS` after a full download. Release tags identify immutable intended snapshots; pin the tag or commit for experiments. Later assets should extend task-specific paths and receive a new version and changelog entry.

SHA256SUMS

@@ -0,0 +1,31 @@
+72da23bf0d02b4c74bdb144ff1e2dc808a406a8f95a509910c4bd8d41c1045aa  CHANGELOG.md
+ae579a4341144a2a2e4d8d629df796e113a877e077eca049383c170a70d9a0e1  README.md
+8e8ffccb15a01db37b5e5db7c9bfbdc6189b54350580924b1e97865225c8effd  level1/README.md
+bfef43e1dc61d0bdd984300809a648b5a867dd91d88752f335ed25bea8b90cd7  level1/audio/motion_instructions/data.parquet
+5df4702911e2e99ec798e440fd40a01d2e08260fd01a9ce9a5c83b7f279d3e1f  level1/image/img_txt_human/index.parquet
+57bf240f287542fcd6f791d28f247a0b4792e39fb195a5aedfbf8e15f11dee34  level1/image/img_txt_human/videos-00000-of-00005.index.parquet
+6dc5bc4ff83d2af9399f69ab690a52739b835f1b9f53bb1f17f8d8512fb0bd75  level1/image/img_txt_human/videos-00000-of-00005.tar
+dd9e97780d72808c5b3c3f85fdd9cff9773ba57d24fab0a7ea9ceb7c6d98335b  level1/image/img_txt_human/videos-00001-of-00005.index.parquet
+1cb2c9e291fcd28658ad976a6897d4bd9555ad7578bb48aac6bcefb38bc0535e  level1/image/img_txt_human/videos-00001-of-00005.tar
+2e782fbaa534c9e1069523f40ceaaa63bc7e3493784e7c772fc93164b62cb976  level1/image/img_txt_human/videos-00002-of-00005.index.parquet
+de33cb6e65c0491b6d172ce855c10552c9984bb7832bdeef91d340ddfdf38cf8  level1/image/img_txt_human/videos-00002-of-00005.tar
+6cef1a5344b7072266b48fc66a067ac516184a5d6b88ad6542ef039611d495db  level1/image/img_txt_human/videos-00003-of-00005.index.parquet
+fdda7d190a419e82335e2d22d08fe55c498be2310e951ee3e3d378ac2f0a6c01  level1/image/img_txt_human/videos-00003-of-00005.tar
+c7e0b8a83d8369b17068974765a9f4cb4ca7519cd876100eee7f4d121dd490c9  level1/image/img_txt_human/videos-00004-of-00005.index.parquet
+ae144351bb309ad0e73d97b488ee50da6e79cf5aab99474da665ba163f204eb4  level1/image/img_txt_human/videos-00004-of-00005.tar
+201017542e301ebb04056ec8373e7f6cdef5a1f07a3bb420b313432c14525d8f  level1/image/img_txt_skel/index.parquet
+201017542e301ebb04056ec8373e7f6cdef5a1f07a3bb420b313432c14525d8f  level1/image/img_txt_skel/videos-00000-of-00001.index.parquet
+1dfa5fadb895061f3a3581db68e50b7ac9890de43343e89e3e183098e8977849  level1/image/img_txt_skel/videos-00000-of-00001.tar
+ee5b35f9815831d734933dd35978069203d629269d3b9f5db4a80415d374d83e  level1/manifests/assets.json
+25336b3f568fa1c8e7aa4f8915193ea52b0e385e37be7e3f2f2d6fb811301456  level1/manifests/provenance.json
+68729ecef845bc4128b1a4eea711c69dea604f6dff39af7e93749034e712dede  level1/manifests/validation.json
+3fe0a5bf1a6864425a1dc67c9428c1c4a583b68ef8cb26a83538ded8ae082754  level1/text/img_txt/data.parquet
+cecc1813667fc4924a462818985f29c9f976448ddad071b1b4a557d3e037ae37  level1/text/mul_bal_txt/data.parquet
+ca153ad67cd61685623a89d6a32e68cfb0fccec0e0a9a45c9c970880016ae936  level1/text/txt_imit/txt_comp_hands/data.parquet
+67f54b5974dce81d59237b2056c6ac3439375be13ef9277d4c501abf29a8597e  level1/text/txt_imit/txt_comp_legs/data.parquet
+b78ce4e03cd89487588edd03626554269c1d2fbd8199f268475ab5bc503fa27a  level1/text/txt_imit/txt_gen/data.parquet
+0ff3312de5dec97154d2cbf0841b34bf665b6a7c95d128d0dcf039301f415d28  level1/text/txt_pred/txt_fore/data.parquet
+bfcb12d7609406aeb55ccecb0b79762c9d228478069a6dde15cd7779049031b7  level1/text/txt_pred/txt_inter/data.parquet
+4b3085f251d45b94a1eed51760393477b74b81964d168546020f546c2f1606d7  level1/text/txt_pred/txt_retro/data.parquet
+f831c77f531e6bc3f3144ae7a77f22804e313853a9647c1d214c474674d777ef  scripts/build_dataset.py
+5be165f88269e4fecc76a8a76bedc1b423488704897c4afb656eae8c766845c3  scripts/restore_text.py

level1/README.md

@@ -0,0 +1,21 @@
+# Level 1 预处理中间资产
+
+本发布目录独立于工作目录，包含本次选定的整理后文本和 IMG_TXT 条件视频。
+
+| 文本配置 | 条数 | 仓库路径 |
+|---|---:|---|
+| level1_txt_gen | 38,522 | `level1/text/txt_imit/txt_gen/data.parquet` |
+| level1_txt_comp_hands | 38,522 | `level1/text/txt_imit/txt_comp_hands/data.parquet` |
+| level1_txt_comp_legs | 38,522 | `level1/text/txt_imit/txt_comp_legs/data.parquet` |
+| level1_txt_fore | 14,333 | `level1/text/txt_pred/txt_fore/data.parquet` |
+| level1_txt_retro | 14,333 | `level1/text/txt_pred/txt_retro/data.parquet` |
+| level1_txt_inter | 7,764 | `level1/text/txt_pred/txt_inter/data.parquet` |
+| level1_img_txt | 14,333 | `level1/text/img_txt/data.parquet` |
+| level1_mul_bal_txt | 8,193 | `level1/text/mul_bal_txt/data.parquet` |
+| level1_audio_motion_instructions | 38,522 | `level1/audio/motion_instructions/data.parquet` |
+
+HUMAN 与 SKEL 各 14,333 个 MP4，保存在 image/ 下独立 tar 分片中；index.parquet 可定位每个视频及 JSON 元数据。已有 JSON 中的服务器绝对图片路径已改为相对于外部 public 根的路径。
+
+音频类仅包含整理后的动作指令；本发布不含音频实体或转写全文。MUL_BAL 目前仅包含文本。各配置的 data 表示完整资产集合，不代表训练集划分。
+
+详见仓库根 README、manifests/ 清单和 SHA256SUMS。

level1/audio/motion_instructions/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfef43e1dc61d0bdd984300809a648b5a867dd91d88752f335ed25bea8b90cd7
+size 6600405

level1/image/img_txt_human/index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5df4702911e2e99ec798e440fd40a01d2e08260fd01a9ce9a5c83b7f279d3e1f
+size 3278467

level1/image/img_txt_human/videos-00000-of-00005.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57bf240f287542fcd6f791d28f247a0b4792e39fb195a5aedfbf8e15f11dee34
+size 704524

level1/image/img_txt_human/videos-00000-of-00005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dc5bc4ff83d2af9399f69ab690a52739b835f1b9f53bb1f17f8d8512fb0bd75
+size 1053890560

level1/image/img_txt_human/videos-00001-of-00005.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd9e97780d72808c5b3c3f85fdd9cff9773ba57d24fab0a7ea9ceb7c6d98335b
+size 700675

level1/image/img_txt_human/videos-00001-of-00005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cb2c9e291fcd28658ad976a6897d4bd9555ad7578bb48aac6bcefb38bc0535e
+size 1054064640

level1/image/img_txt_human/videos-00002-of-00005.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e782fbaa534c9e1069523f40ceaaa63bc7e3493784e7c772fc93164b62cb976
+size 708865

level1/image/img_txt_human/videos-00002-of-00005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de33cb6e65c0491b6d172ce855c10552c9984bb7832bdeef91d340ddfdf38cf8
+size 1053849600

level1/image/img_txt_human/videos-00003-of-00005.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cef1a5344b7072266b48fc66a067ac516184a5d6b88ad6542ef039611d495db
+size 676290

level1/image/img_txt_human/videos-00003-of-00005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdda7d190a419e82335e2d22d08fe55c498be2310e951ee3e3d378ac2f0a6c01
+size 1054638080

level1/image/img_txt_human/videos-00004-of-00005.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7e0b8a83d8369b17068974765a9f4cb4ca7519cd876100eee7f4d121dd490c9
+size 506480

level1/image/img_txt_human/videos-00004-of-00005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae144351bb309ad0e73d97b488ee50da6e79cf5aab99474da665ba163f204eb4
+size 743598080

level1/image/img_txt_skel/index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:201017542e301ebb04056ec8373e7f6cdef5a1f07a3bb420b313432c14525d8f
+size 3252270

level1/image/img_txt_skel/videos-00000-of-00001.index.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:201017542e301ebb04056ec8373e7f6cdef5a1f07a3bb420b313432c14525d8f
+size 3252270

level1/image/img_txt_skel/videos-00000-of-00001.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dfa5fadb895061f3a3581db68e50b7ac9890de43343e89e3e183098e8977849
+size 308602880

level1/manifests/assets.json

@@ -0,0 +1,84 @@
+{
+  "version": "v1.0.0",
+  "text": [
+    {
+      "config": "level1_txt_gen",
+      "task": "TXT_GEN",
+      "path": "level1/text/txt_imit/txt_gen/data.parquet",
+      "records": 38522,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_txt_comp_hands",
+      "task": "TXT_COMP_HANDS",
+      "path": "level1/text/txt_imit/txt_comp_hands/data.parquet",
+      "records": 38522,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_txt_comp_legs",
+      "task": "TXT_COMP_LEGS",
+      "path": "level1/text/txt_imit/txt_comp_legs/data.parquet",
+      "records": 38522,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_txt_fore",
+      "task": "TXT_FORE",
+      "path": "level1/text/txt_pred/txt_fore/data.parquet",
+      "records": 14333,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_txt_retro",
+      "task": "TXT_RETRO",
+      "path": "level1/text/txt_pred/txt_retro/data.parquet",
+      "records": 14333,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_txt_inter",
+      "task": "TXT_INTER",
+      "path": "level1/text/txt_pred/txt_inter/data.parquet",
+      "records": 7764,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_img_txt",
+      "task": "IMG_TXT_SKEL",
+      "path": "level1/text/img_txt/data.parquet",
+      "records": 14333,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_mul_bal_txt",
+      "task": "MUL_BAL",
+      "path": "level1/text/mul_bal_txt/data.parquet",
+      "records": 8193,
+      "empty_text_records": 0
+    },
+    {
+      "config": "level1_audio_motion_instructions",
+      "task": "AUDIO_MOTION_INSTRUCTIONS",
+      "path": "level1/audio/motion_instructions/data.parquet",
+      "records": 38522,
+      "empty_text_records": 0
+    }
+  ],
+  "videos": [
+    {
+      "task": "IMG_TXT_HUMAN",
+      "path": "level1/image/img_txt_human",
+      "records": 14333,
+      "shards": 5,
+      "source_mp4_bytes": 4930718649
+    },
+    {
+      "task": "IMG_TXT_SKEL",
+      "path": "level1/image/img_txt_skel",
+      "records": 14333,
+      "shards": 1,
+      "source_mp4_bytes": 279026980
+    }
+  ]
+}

level1/manifests/provenance.json

@@ -0,0 +1,17 @@
+{
+  "release": "v1.0.0",
+  "packaged_date": "2026-09-17",
+  "packaging_script_sha256": "f831c77f531e6bc3f3144ae7a77f22804e313853a9647c1d214c474674d777ef",
+  "pyarrow_version": "24.0.0",
+  "text_policy": "strict UTF-8; bytes and line endings preserved",
+  "mp4_policy": "unmodified source bytes",
+  "metadata_policy": "external image paths normalized to public-root-relative",
+  "shard_target_bytes": 1073741824,
+  "tar_policy": "independent PAX tar, sorted filenames, normalized headers",
+  "source_generator_fingerprints": {
+    "build_keyframe_static_videos.py": "8a74e03bfd4b4481f37241e4aac21eafd12e75470fb73917071330e85f469a2a",
+    "run_img_txt_video_to_motion_fast.sh": "9b5a39932dc5314f2d24a660223a15501c6c9575834758d0bce23c5bb58110f2"
+  },
+  "source_generation_commit": null,
+  "source_generation_commit_note": "Historical generation commit was not recorded"
+}

level1/manifests/validation.json

@@ -0,0 +1,11 @@
+{
+  "text_records": 213044,
+  "video_records": 28666,
+  "text_utf8_roundtrip_sha256": "passed for every row",
+  "video_and_metadata_tar_payload_sha256": "passed for every member",
+  "duplicate_asset_ids": 0,
+  "human_skel_matching_source_ids": 14333,
+  "img_text_skel_matching_source_ids": 14333,
+  "empty_text_records": 0,
+  "mp4_decoding_test": "not performed; lossless packaging only"
+}

level1/text/img_txt/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fe0a5bf1a6864425a1dc67c9428c1c4a583b68ef8cb26a83538ded8ae082754
+size 1183629

level1/text/mul_bal_txt/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cecc1813667fc4924a462818985f29c9f976448ddad071b1b4a557d3e037ae37
+size 855141

level1/text/txt_imit/txt_comp_hands/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca153ad67cd61685623a89d6a32e68cfb0fccec0e0a9a45c9c970880016ae936
+size 5063421

level1/text/txt_imit/txt_comp_legs/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67f54b5974dce81d59237b2056c6ac3439375be13ef9277d4c501abf29a8597e
+size 5509060

level1/text/txt_imit/txt_gen/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b78ce4e03cd89487588edd03626554269c1d2fbd8199f268475ab5bc503fa27a
+size 6558848

level1/text/txt_pred/txt_fore/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ff3312de5dec97154d2cbf0841b34bf665b6a7c95d128d0dcf039301f415d28
+size 2090675

level1/text/txt_pred/txt_inter/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfcb12d7609406aeb55ccecb0b79762c9d228478069a6dde15cd7779049031b7
+size 1155631

level1/text/txt_pred/txt_retro/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b3085f251d45b94a1eed51760393477b74b81964d168546020f546c2f1606d7
+size 2072849

scripts/build_dataset.py

@@ -0,0 +1,371 @@
+#!/usr/bin/env python3
+"""Build the RoboSteer Level 1 publication tree without modifying source assets."""
+import argparse
+import concurrent.futures
+import hashlib
+import io
+import json
+import os
+from pathlib import Path
+import shutil
+import tarfile
+import time
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+VERSION = "v1.0.0"
+TEXT = [
+    ("txt_gen", "text/processed/TXT/IMIT/TXT_GEN", "text/txt_imit/txt_gen", "TXT_GEN", 38522),
+    ("txt_comp_hands", "text/processed/TXT/IMIT/TXT_COMP_HANDS", "text/txt_imit/txt_comp_hands", "TXT_COMP_HANDS", 38522),
+    ("txt_comp_legs", "text/processed/TXT/IMIT/TXT_COMP_LEGS", "text/txt_imit/txt_comp_legs", "TXT_COMP_LEGS", 38522),
+    ("txt_fore", "text/processed/TXT/PRED/TXT_FORE", "text/txt_pred/txt_fore", "TXT_FORE", 14333),
+    ("txt_retro", "text/processed/TXT/PRED/TXT_RETRO", "text/txt_pred/txt_retro", "TXT_RETRO", 14333),
+    ("txt_inter", "text/processed/TXT/PRED/TXT_INTER", "text/txt_pred/txt_inter", "TXT_INTER", 7764),
+    ("img_txt", "text/processed/IMG_TXT", "text/img_txt", "IMG_TXT_SKEL", 14333),
+    ("mul_bal_txt", "text/processed/MUL_BAL_TXT", "text/mul_bal_txt", "MUL_BAL", 8193),
+    ("audio_motion_instructions", "audio/processed/whisper_txt_preprocessed", "audio/motion_instructions", "AUDIO_MOTION_INSTRUCTIONS", 38522),
+]
+
+
+def log(*args):
+    print(time.strftime("%Y-%m-%d %H:%M:%S"), *args, flush=True)
+
+
+def digest(path):
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        for block in iter(lambda: f.read(4 * 1024 * 1024), b""):
+            h.update(block)
+    return h.hexdigest()
+
+
+def write_json(path, value):
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+
+
+def write_parquet(path, rows):
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(".parquet.tmp")
+    pq.write_table(pa.Table.from_pylist(rows), tmp, compression="zstd", row_group_size=10000)
+    tmp.replace(path)
+
+
+def source_id(stem, task):
+    prefix = "L1_" + task + "_"
+    if stem.startswith(prefix):
+        return stem[len(prefix):]
+    if task == "AUDIO_MOTION_INSTRUCTIONS" and stem.endswith("_audio"):
+        return stem[:-len("_audio")]
+    raise ValueError(f"Unrecognized sample naming: {task}: {stem}")
+
+
+def read_unchanged(path):
+    before = path.stat()
+    data = path.read_bytes()
+    after = path.stat()
+    assert (before.st_size, before.st_mtime_ns) == (after.st_size, after.st_mtime_ns), path
+    return data
+
+
+def build_text(source, release):
+    manifest = []
+    ids = {}
+    for config, src, dst, task, expected in TEXT:
+        files = sorted((source / src).rglob("*.txt"))
+        assert len(files) == expected, (src, len(files), expected)
+
+        def convert(path):
+            assert not path.is_symlink(), path
+            data = read_unchanged(path)
+            return {
+                "sample_id": path.stem,
+                "source_sample_id": source_id(path.stem, task),
+                "task": task,
+                "level": "level1",
+                "source_modality": "audio" if config.startswith("audio_") else "text",
+                "text": data.decode("utf-8"),
+                "original_relpath": str(path.relative_to(source)),
+                "size_bytes": len(data),
+                "sha256": hashlib.sha256(data).hexdigest(),
+            }
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=12) as pool:
+            rows = list(pool.map(convert, files))
+        assert len({x["sample_id"] for x in rows}) == expected
+        dest = release / "level1" / dst / "data.parquet"
+        write_parquet(dest, rows)
+        restored = pq.read_table(dest).to_pylist()
+        assert len(restored) == expected
+        for row in restored:
+            raw = row["text"].encode("utf-8")
+            assert len(raw) == row["size_bytes"]
+            assert hashlib.sha256(raw).hexdigest() == row["sha256"]
+        ids[config] = {x["source_sample_id"] for x in rows}
+        manifest.append({"config": "level1_" + config, "task": task,
+                         "path": str(dest.relative_to(release)), "records": expected,
+                         "empty_text_records": sum(not x["text"].strip() for x in rows)})
+        log("TEXT_VERIFIED", config, expected)
+    return manifest, ids
+
+
+class HashReader:
+    def __init__(self, f):
+        self.f = f
+        self.hash = hashlib.sha256()
+
+    def read(self, n=-1):
+        data = self.f.read(n)
+        self.hash.update(data)
+        return data
+
+
+def public_relative(value):
+    if "/public/" in value:
+        value = value.split("/public/", 1)[1]
+    p = Path(value)
+    if p.is_absolute() or ".." in p.parts:
+        raise ValueError("Image provenance must be relative to public root")
+    return p.as_posix()
+
+
+def tar_info(name, size):
+    entry = tarfile.TarInfo(name)
+    entry.size = size
+    entry.mode = 0o644
+    entry.uid = entry.gid = entry.mtime = 0
+    return entry
+
+
+def build_video(source, release, task, target_bytes):
+    folder = release / "level1/image" / task.lower()
+    folder.mkdir(parents=True, exist_ok=True)
+    source_dir = source / "img+mul_bal/output" / task
+    metadata_dir = source / "img+mul_bal/inputs/keyframe_static_videos" / task
+    videos = sorted(source_dir.glob("*.mp4"))
+    assert len(videos) == 14333
+    assert {p.stem for p in videos} == {p.stem for p in metadata_dir.glob("*.json")}
+    groups, group, size = [], [], 0
+    for p in videos:
+        estimate = ((p.stat().st_size + 511) // 512) * 512 + 8192
+        if group and size + estimate > target_bytes:
+            groups.append(group)
+            group, size = [], 0
+        group.append(p)
+        size += estimate
+    if group:
+        groups.append(group)
+    all_rows = []
+    for i, members in enumerate(groups):
+        shard = folder / f"videos-{i:05d}-of-{len(groups):05d}.tar"
+        checkpoint = shard.with_suffix(".index.parquet")
+        if shard.exists() and checkpoint.exists():
+            rows = pq.read_table(checkpoint).to_pylist()
+            assert [x["sample_id"] for x in rows] == [p.stem for p in members]
+            all_rows.extend(rows)
+            log("REUSE_SHARD", shard.name)
+            continue
+        rows = []
+        temp = shard.with_suffix(".tar.tmp")
+        with tarfile.open(temp, "w", format=tarfile.PAX_FORMAT) as tar:
+            for p in members:
+                assert not p.is_symlink()
+                before = p.stat()
+                meta_path = metadata_dir / (p.stem + ".json")
+                raw_meta = read_unchanged(meta_path)
+                meta = json.loads(raw_meta)
+                assert meta["task_id"] == p.stem
+                meta["input_images"] = [public_relative(x) for x in meta["input_images"]]
+                meta["input_images_base"] = "external_public_root"
+                normalized = (json.dumps(meta, ensure_ascii=False, indent=2) + "\n").encode("utf-8")
+                with p.open("rb") as f:
+                    reader = HashReader(f)
+                    tar.addfile(tar_info(p.name, before.st_size), reader)
+                after = p.stat()
+                assert (before.st_size, before.st_mtime_ns) == (after.st_size, after.st_mtime_ns)
+                tar.addfile(tar_info(p.stem + ".json", len(normalized)), io.BytesIO(normalized))
+                rows.append({
+                    "sample_id": p.stem, "source_sample_id": source_id(p.stem, task),
+                    "task": task, "level": "level1", "source_modality": "image",
+                    "shard": str(shard.relative_to(release)), "member_path": p.name,
+                    "metadata_member_path": p.stem + ".json",
+                    "original_relpath": str(p.relative_to(source)),
+                    "size_bytes": before.st_size, "sha256": reader.hash.hexdigest(),
+                    "metadata_sha256": hashlib.sha256(normalized).hexdigest(),
+                    "source_metadata_sha256": hashlib.sha256(raw_meta).hexdigest(),
+                    "text": meta["text"],
+                    "source_duration_seconds": float(meta["source_duration_seconds"]),
+                    "generated_duration_seconds": float(meta["generated_duration_seconds"]),
+                    "timestamps_seconds": meta["timestamps_seconds"],
+                    "static_frame_durations_seconds": meta["static_frame_durations_seconds"],
+                    "input_images": meta["input_images"],
+                })
+        temp.replace(shard)
+        write_parquet(checkpoint, rows)
+        all_rows.extend(rows)
+        log("SHARD_WRITTEN", task, shard.name, len(rows), shard.stat().st_size)
+    assert len(all_rows) == 14333
+    assert len({x["sample_id"] for x in all_rows}) == 14333
+    write_parquet(folder / "index.parquet", all_rows)
+    # Read all tar payloads back, checking hashes and exact archive membership.
+    for shard in sorted(folder.glob("*.tar")):
+        wanted = {}
+        for row in all_rows:
+            if row["shard"] == str(shard.relative_to(release)):
+                wanted[row["member_path"]] = row["sha256"]
+                wanted[row["metadata_member_path"]] = row["metadata_sha256"]
+        seen = set()
+        with tarfile.open(shard, "r|") as tar:
+            for entry in tar:
+                assert entry.isfile() and entry.name in wanted and entry.name not in seen
+                f = tar.extractfile(entry)
+                h = hashlib.sha256()
+                for block in iter(lambda: f.read(1024 * 1024), b""):
+                    h.update(block)
+                assert h.hexdigest() == wanted[entry.name]
+                seen.add(entry.name)
+        assert seen == set(wanted)
+        log("SHARD_VERIFIED", task, shard.name)
+    # Per-shard indexes are kept: they support selective download and resumable builds.
+    return {"task": task, "path": str(folder.relative_to(release)), "records": len(all_rows),
+            "shards": len(groups), "source_mp4_bytes": sum(x["size_bytes"] for x in all_rows)}, {
+                x["source_sample_id"] for x in all_rows}
+
+
+def build_docs(release, texts, videos, validation):
+    config = []
+    for row in texts:
+        config += [f"- config_name: {row['config']}", "  data_files:", "  - split: data",
+                   f"    path: {row['path']}"]
+    for row in videos:
+        config += [f"- config_name: level1_{row['task'].lower()}_index", "  data_files:",
+                   "  - split: data", f"    path: {row['path']}/index.parquet"]
+    readme = "---\npretty_name: RoboSteer Preprocessing\nlanguage:\n- en\ntags:\n- robosteer\n- preprocessing\n- motion-generation\n- intermediate-assets\nconfigs:\n" + "\n".join(config) + "\n---\n\n"
+    readme += """# RoboSteer Preprocessing
+
+Reusable intermediate preprocessing assets for RoboSteer. This initial release contains Level 1 processed text instructions and image-conditioned static videos. Other levels and preprocessing stages can be added under their own directories.
+
+## Release v1.0.0
+
+- **213,044 text records** in nine lossless UTF-8 Parquet tables.
+- **28,666 original MP4 files**: 14,333 IMG_TXT_HUMAN and 14,333 IMG_TXT_SKEL.
+- Independent, uncompressed tar shards contain MP4 files and their normalized JSON metadata.
+- Per-sample indexes, release inventory, SHA-256 checksums, and validation report.
+
+These are preprocessing assets, not a declared train/validation/test split. The `data` split means the complete asset collection for a configuration. Audio motion instructions are text derived from audio; audio recordings and raw transcripts are not included. MUL_BAL is included as processed text; this release does not contain MUL_BAL videos. Prediction-video manifests are outside this release scope.
+
+## Layout
+
+```text
+level1/
+  text/txt_imit/{txt_gen,txt_comp_hands,txt_comp_legs}/data.parquet
+  text/txt_pred/{txt_fore,txt_retro,txt_inter}/data.parquet
+  text/{img_txt,mul_bal_txt}/data.parquet
+  audio/motion_instructions/data.parquet
+  image/{img_txt_human,img_txt_skel}/
+    videos-00000-of-NNNNN.tar
+    videos-00000-of-NNNNN.index.parquet
+    index.parquet
+  manifests/{assets,validation,provenance}.json
+scripts/restore_text.py
+scripts/build_dataset.py
+SHA256SUMS
+CHANGELOG.md
+```
+
+## Load text or an index
+
+```python
+from datasets import load_dataset
+ds = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_txt_gen", split="data", revision="v1.0.0")
+video_index = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_img_txt_human_index", split="data", revision="v1.0.0")
+```
+
+Text columns: `sample_id`, `source_sample_id`, `task`, `level`, `source_modality`, `text`, `original_relpath`, `size_bytes`, `sha256`. Original UTF-8 bytes, including line endings, can be reconstructed with `text.encode('utf-8')`. Use `scripts/restore_text.py` to restore the original relative file layout.
+
+## Download video shards
+
+```python
+from huggingface_hub import snapshot_download
+snapshot_download("PhoebeCC/RoboSteer-Preprocessing", repo_type="dataset",
+    revision="v1.0.0", local_dir="robosteer-preprocessing",
+    allow_patterns=["level1/image/img_txt_human/*", "level1/manifests/*", "README.md", "SHA256SUMS"])
+```
+
+Each tar is independently extractable. The index identifies its repository-relative `shard`, MP4 `member_path`, JSON `metadata_member_path`, and checksums. Video indexes are searchable tables; this release does not configure an embedded-video viewer. Extract each task into a separate folder. MP4 files are preserved byte for byte, without re-encoding.
+
+## Provenance and alignment
+
+`sample_id` preserves the source filename stem. `source_sample_id` removes only the known task prefix (or audio suffix). Use `(task, sample_id)` as an asset key; use `source_sample_id` to join related tasks, without assuming all tasks cover the same samples. Coverage results are recorded in the validation report.
+
+Video metadata originates from existing sidecar JSON files. Metadata text is the original sidecar description, not necessarily the rewritten `text/img_txt` instruction. Image paths are converted to paths relative to the external public-data root. Source duration and generated duration are metadata values, not newly measured properties. The source generator specifies 25 FPS; this packaging run does not probe actual MP4 frame rates. Source images, external task JSONs, model weights, raw audio, audit logs, and organization backups are not bundled.
+
+## Rights and citation
+
+The repository owner has not yet supplied a release license or complete upstream attribution. No blanket open-source license is asserted for these derived assets. Upstream restrictions remain applicable; consult the owner about permitted redistribution and use. Formal benchmark citation and upstream dataset attribution will be added when supplied.
+
+## Reproducibility
+
+See `level1/manifests/provenance.json` for packaging rules and generator fingerprints, `assets.json` for counts, and `validation.json` for checks. `SHA256SUMS` covers all payload and documentation files except itself. SHA-256 verification can be performed with `sha256sum -c SHA256SUMS` after a full download. Release tags identify immutable intended snapshots; pin the tag or commit for experiments. Later assets should extend task-specific paths and receive a new version and changelog entry.
+"""
+    (release / "README.md").write_text(readme)
+    (release / "CHANGELOG.md").write_text("# Changelog\n\n## v1.0.0 — 2026-09-17\n\nInitial Level 1 release: nine processed-text tables, two static-video tasks, original-name recovery, metadata indexes, and SHA-256 verification. No source text rewriting or video re-encoding.\n")
+    table = "\n".join(f"| {x['config']} | {x['records']:,} | `{x['path']}` |" for x in texts)
+    (release / "level1/README.md").write_text("# Level 1 预处理中间资产\n\n本发布目录独立于工作目录，包含本次选定的整理后文本和 IMG_TXT 条件视频。\n\n| 文本配置 | 条数 | 仓库路径 |\n|---|---:|---|\n" + table + "\n\nHUMAN 与 SKEL 各 14,333 个 MP4，保存在 image/ 下独立 tar 分片中；index.parquet 可定位每个视频及 JSON 元数据。已有 JSON 中的服务器绝对图片路径已改为相对于外部 public 根的路径。\n\n音频类仅包含整理后的动作指令；本发布不含音频实体或转写全文。MUL_BAL 目前仅包含文本。各配置的 data 表示完整资产集合，不代表训练集划分。\n\n详见仓库根 README、manifests/ 清单和 SHA256SUMS。\n")
+    write_json(release / "level1/manifests/assets.json", {"version": VERSION, "text": texts, "videos": videos})
+    write_json(release / "level1/manifests/validation.json", validation)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--source", type=Path, required=True, help="Current level1 work directory")
+    parser.add_argument("--release", type=Path, required=True, help="Dedicated publication directory")
+    parser.add_argument("--shard-bytes", type=int, default=1024**3)
+    args = parser.parse_args()
+    assert args.source.resolve() != args.release.resolve()
+    args.release.mkdir(parents=True, exist_ok=True)
+    text, ids = build_text(args.source, args.release)
+    videos = []
+    for task in ["IMG_TXT_HUMAN", "IMG_TXT_SKEL"]:
+        row, sample_ids = build_video(args.source, args.release, task, args.shard_bytes)
+        videos.append(row)
+        ids[task] = sample_ids
+    validation = {
+        "text_records": sum(x["records"] for x in text), "video_records": sum(x["records"] for x in videos),
+        "text_utf8_roundtrip_sha256": "passed for every row",
+        "video_and_metadata_tar_payload_sha256": "passed for every member",
+        "duplicate_asset_ids": 0,
+        "human_skel_matching_source_ids": len(ids["IMG_TXT_HUMAN"] & ids["IMG_TXT_SKEL"]),
+        "img_text_skel_matching_source_ids": len(ids["img_txt"] & ids["IMG_TXT_SKEL"]),
+        "empty_text_records": sum(x["empty_text_records"] for x in text),
+        "mp4_decoding_test": "not performed; lossless packaging only",
+    }
+    assert validation["text_records"] == 213044
+    assert validation["human_skel_matching_source_ids"] == 14333
+    assert validation["img_text_skel_matching_source_ids"] == 14333
+    build_docs(args.release, text, videos, validation)
+    scripts = args.release / "scripts"
+    scripts.mkdir(exist_ok=True)
+    shutil.copy2(__file__, scripts / "build_dataset.py")
+    shutil.copy2(Path(__file__).with_name("restore_text.py"), scripts / "restore_text.py")
+    generators = [args.source / "img+mul_bal/scripts/build_keyframe_static_videos.py",
+                  args.source.parent / "GEM/scripts/demo/run_img_txt_video_to_motion_fast.sh"]
+    write_json(args.release / "level1/manifests/provenance.json", {
+        "release": VERSION, "packaged_date": "2026-09-17", "packaging_script_sha256": digest(Path(__file__)),
+        "pyarrow_version": pa.__version__, "text_policy": "strict UTF-8; bytes and line endings preserved",
+        "mp4_policy": "unmodified source bytes", "metadata_policy": "external image paths normalized to public-root-relative",
+        "shard_target_bytes": args.shard_bytes, "tar_policy": "independent PAX tar, sorted filenames, normalized headers",
+        "source_generator_fingerprints": {p.name: digest(p) for p in generators},
+        "source_generation_commit": None, "source_generation_commit_note": "Historical generation commit was not recorded",
+    })
+    checks = []
+    for p in sorted(args.release.rglob("*")):
+        if p.is_file() and p.name != "SHA256SUMS" and ".cache" not in p.parts:
+            checks.append(f"{digest(p)}  {p.relative_to(args.release).as_posix()}")
+    (args.release / "SHA256SUMS").write_text("\n".join(checks) + "\n")
+    log("BUILD_COMPLETE", json.dumps(validation))
+
+
+if __name__ == "__main__":
+    main()

scripts/restore_text.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""Restore text Parquet assets to their original relative paths, checking hashes."""
+import argparse
+import hashlib
+from pathlib import Path, PurePosixPath
+import pyarrow.parquet as pq
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--dataset-root", type=Path, required=True)
+    p.add_argument("--output", type=Path, required=True)
+    args = p.parse_args()
+    count = 0
+    for base in ["level1/text", "level1/audio"]:
+        for parquet in sorted((args.dataset_root / base).rglob("*.parquet")):
+            for batch in pq.ParquetFile(parquet).iter_batches(batch_size=4096):
+                for row in batch.to_pylist():
+                    rel = PurePosixPath(row["original_relpath"])
+                    if rel.is_absolute() or ".." in rel.parts:
+                        raise ValueError("Unsafe relative path")
+                    raw = row["text"].encode("utf-8")
+                    assert hashlib.sha256(raw).hexdigest() == row["sha256"]
+                    out = args.output / str(rel)
+                    # Refuse writing outside the requested output tree through symlinks.
+                    if not out.resolve().is_relative_to(args.output.resolve()):
+                        raise ValueError("Output path escapes root")
+                    out.parent.mkdir(parents=True, exist_ok=True)
+                    if out.exists():
+                        if out.read_bytes() != raw:
+                            raise FileExistsError(out)
+                    else:
+                        with out.open("xb") as f:
+                            f.write(raw)
+                    count += 1
+    print(f"Restored and verified {count} text files")
+
+
+if __name__ == "__main__":
+    main()