Preserve DOCCI acquisition candidate

README.md

@@ -1,30 +1,11 @@
 ---
 license: cc-by-4.0
-language:
-- en
-task_categories:
-- visual-question-answering
 ---
-# SetGround DOCCI 五图交互匹配任务
+# SetGround DOCCI artifacts
 
-此版本只有任务与首批图片，尚无模型问题、答案、视觉分数、训练模型或准确率。
-来源为Google LLC发布、Onoe等构建的[DOCCI](https://google.github.io/docci/)（ECCV 2024），图像与描述遵守[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)。本派生版本改变了图像分组、描述截取、排列和train内开发划分，并非官方新benchmark。完整原描述保留作来源与监督参考，不能进入部署观察。
+Derived from Google DOCCI (Onoe et al., ECCV 2024), under CC BY 4.0.
 
-## 协议与来源
+- [Fixed tasks and 560 pilot images](snapshots/2026-10-24-task-manifest-v1/README.md).
+- [Fixed-model acquisition recovery snapshots](snapshots/2026-10-24-acquisition-h100-v1). Each stage has a manifest and archive; stage exit codes and per-shard receipts identify completed versus partial data.
 
-保留官方train/test图片边界，qual不使用；每个split/cluster按seed43哈希排序，连续五图成组、余数弃用。train内每cluster留max(1,floor(组数/5))组dev，其余train。候选和目标各自独立哈希排列，actor初始输入为原描述第一个句末标点片段。完整规则见Git `docs/setground_docci_tasks_20261024.md`。
-train=1548、dev=318、test=941组，每组5图；同一源ID/文件名只出现一次。pilot为64/16/32组、560图，按cluster轮流固定采样，不按模型成绩挑选。共享类别和复现实体仍存在，不是独立场景划分；summary报告共享标签。首句可能已经充分辨识，cluster不保证难度，没有过滤已知失败例。
-
-## 文件与权限边界
-
-- actors.jsonl.gz：每行id、split、candidate_images（5个文件名）、initial_descriptions（5行初始描述）。candidate_images在图片archive的images/下；全量图片可从固定上游archive恢复，当前HF只含pilot。
-- oracle.jsonl.gz：同id的assignment（0-based列排列）、candidate/target_example_ids、完整描述、官方split、cluster和entity_tags。只用于来源、训练监督和评测，不供actor、问题生成器或Director读取。
-- pilot.json：固定三split的episode IDs；omitted.jsonl.gz：每个未使用源ID及原因。
-- summary.json：源GCS generation/MD5/SHA、任务数量/共享实体/构建检查、源码和配置；images-summary.json：每图文件SHA、decoded RGB SHA、尺寸、episode/split、archive路径和完整上游图片archive SHA。
-- pilot-images.tar：560张未改动JPEG；精确像素重复检查见images-summary，不等于近重复或物理场景独立。
-
-## 复现
-
-Git <https://github.com/luojiaxuan/cvpr-multimodal-evidence>；任务源码 `30298d1cbe904b3978bc7dded3633ba60706324d`，图片源码 `bd1d8873ae2c95fecb8b28bf0d922b2e94cc6afa`。
-在源码根运行 `python3 code/prepare_setground_docci_tasks.py --config code/configs/setground_docci_tasks.json --output <tasks> --code-revision 30298d1cbe904b3978bc7dded3633ba60706324d`；图片用 `python3 code/prepare_setground_docci_images.py --tasks <tasks> --output <images> --code-revision bd1d8873ae2c95fecb8b28bf0d922b2e94cc6afa`。
-上游描述、metadata及图片的固定URL/generation/大小/MD5均在summary；图片流式读取整个7.59GBarchive但只落盘pilot，不需要把大包放在Mac。可从本HF固定revision直接下载任务和pilot图片。尚未取得新视觉交互证据，不把这份数据准备当作学习或RL效果。
+Model-generated answers are simulated, not human annotations. No learned-policy or RL result is claimed. Use the immutable revision recorded by the Git execution receipt for reproducible recovery.

snapshots/2026-10-24-acquisition-h100-v1/candidate/README.md

@@ -0,0 +1,7 @@
+# DOCCI fixed-model acquisition recovery snapshot
+
+Source: Google DOCCI, CC BY 4.0; source task revision a6507175f04f78d0ef0449bef8dcad0e7f566398. Generated questions and simulated answers are not human annotations or learned-policy results.
+
+The archive contains plan.json, runtime.json, logs, per-episode candidate/director outputs, and per-call evidence and physical-call ledgers. Missing stages remain incomplete. Restore under the same container input/cache paths to preserve resume identities; images and pinned model snapshots are downloaded separately. See manifest.json for member SHA256.
+
+Code: https://github.com/luojiaxuan/cvpr-multimodal-evidence; runtime.json records the runner revision. Generation: run_setground_docci_batch.py --root <persistent-root> --deadline <unix> --revision <git-sha> --prefix <snapshot-prefix>. Four explicit granted GPUs, 16 fixed shards; science configuration and both model revisions are in plan.json.

snapshots/2026-10-24-acquisition-h100-v1/candidate/acquisition.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2f8740a82b837f67de5265a9adcd08bfbbc4329baeb6f27d0942ddf7e0ded6e
+size 75660

snapshots/2026-10-24-acquisition-h100-v1/candidate/manifest.json

@@ -0,0 +1,122 @@
+{
+  "archive_bytes": 75660,
+  "archive_sha256": "e2f8740a82b837f67de5265a9adcd08bfbbc4329baeb6f27d0942ddf7e0ded6e",
+  "files": [
+    {
+      "bytes": 69535,
+      "path": "plan.json",
+      "sha256": "d379fd2921d4ea6ecab323bbd8c640671f79fc7e23158d091e3cc62b8b0a5b85"
+    },
+    {
+      "bytes": 10235,
+      "path": "runtime.json",
+      "sha256": "360e95b5980ad1c505931ed23a44f88919790ba71349417932322595f41fd205"
+    },
+    {
+      "bytes": 3881,
+      "path": "logs/bootstrap-attempt-01.log",
+      "sha256": "16c0cbedeb590be0ff295cfeb90908dad8d3dae3c86c67b57964c5be457ed5c7"
+    },
+    {
+      "bytes": 159,
+      "path": "logs/candidate-00.log",
+      "sha256": "6e2d1712b345b203d442d4066cb8d3a2c54889403dd2c4e9832fb468b8eca53b"
+    },
+    {
+      "bytes": 159,
+      "path": "logs/candidate-01.log",
+      "sha256": "10a3d893622b5bdcf9268038472a2f84b1f130095fe0d026e43db8d2eb9d3b17"
+    },
+    {
+      "bytes": 159,
+      "path": "logs/candidate-02.log",
+      "sha256": "5cc12e74029fc8a719e09548e99f23d62860bc5b535335f33bef992916248b14"
+    },
+    {
+      "bytes": 159,
+      "path": "logs/candidate-03.log",
+      "sha256": "8ddcf424b68cb969c3ab0f0bd5853783ae49a24606a804613e95391e2bca0d20"
+    },
+    {
+      "bytes": 565,
+      "path": "logs/diagnostic-install.log",
+      "sha256": "c75fd3423140c5ea5dc5e7a1eb63974b982c18cf77915159b9154382786f63e7"
+    },
+    {
+      "bytes": 721,
+      "path": "logs/image-pull.log",
+      "sha256": "c8b29bd2c4817c58212df9c19ffa5a945ebad2c5197289aeb5d38eca7f8433b5"
+    },
+    {
+      "bytes": 13084,
+      "path": "logs/install.log",
+      "sha256": "5e7df025e2272ad5bb0e84512664d9152be38c16c8e4bcda428e268b69a53568"
+    },
+    {
+      "bytes": 0,
+      "path": "logs/patch-backend.log",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "bytes": 159,
+      "path": "logs/profile-forward.log",
+      "sha256": "e3e991b9084cd825a3c3ac2f0f40ceaa19c97206683677cc54a635998c570578"
+    },
+    {
+      "bytes": 11129,
+      "path": "logs/trace-forward-02.log",
+      "sha256": "99ac7b5b9ea3665279b1b849bb8e2faeecc3899f648a8221a3a09b487e5bb9f4"
+    },
+    {
+      "bytes": 10527,
+      "path": "logs/trace-forward.log",
+      "sha256": "d071d3f838bcd39cdab1499fd57b3f81e94b647188af06fecf46ebb9495b1a1c"
+    },
+    {
+      "bytes": 122,
+      "path": "outputs/candidate-batch.json",
+      "sha256": "324ebc93e0232ffcdf8decf97b883464580529c8c330c6b215693322cdbc1715"
+    },
+    {
+      "bytes": 39426,
+      "path": "outputs/ledgers/candidate/00/evidence.jsonl",
+      "sha256": "edf4d3df2a7a3c9f7f2e491905de69eb7a08c7867f077a6be78ef1844750f68d"
+    },
+    {
+      "bytes": 19112,
+      "path": "outputs/ledgers/candidate/00/physical-calls.jsonl",
+      "sha256": "a9a40cc54bfba3318972fe90b4f03575f7be446b8cad95eace05547e08ab97c0"
+    },
+    {
+      "bytes": 63821,
+      "path": "outputs/ledgers/candidate/01/evidence.jsonl",
+      "sha256": "cd85a6cef14c0009bf4796204fc2e223be5fdd88e688e29dfd022f83cf57ccb8"
+    },
+    {
+      "bytes": 29736,
+      "path": "outputs/ledgers/candidate/01/physical-calls.jsonl",
+      "sha256": "fc7c12babf77da7c3fde7d4f031371e28308a81bb47afbb5be3f9f034165fbb0"
+    },
+    {
+      "bytes": 69683,
+      "path": "outputs/ledgers/candidate/02/evidence.jsonl",
+      "sha256": "19499719349e0b5bc642004cf055c4fb3c4fd0d45ec1c0b6043d73be91596648"
+    },
+    {
+      "bytes": 32015,
+      "path": "outputs/ledgers/candidate/02/physical-calls.jsonl",
+      "sha256": "6fc44bc5d945800d99e618c001b746673ed9888f882e0b6fce8ca45022c1575a"
+    },
+    {
+      "bytes": 57376,
+      "path": "outputs/ledgers/candidate/03/evidence.jsonl",
+      "sha256": "1fdabe6620b6aba10ab1f5795965f0de3325adbf773ca6ce04e8e21ba69eb214"
+    },
+    {
+      "bytes": 26920,
+      "path": "outputs/ledgers/candidate/03/physical-calls.jsonl",
+      "sha256": "cfcdf4dc1bffc40f46b4b10d3f139ece3fe1c9ae4bdc2c97fac4c146a483f82f"
+    }
+  ],
+  "stage": "candidate"
+}