Finalize task-transfer catalogue and preserved dataset selections

README.md

@@ -3893,7 +3893,6 @@ Training data, episode-level evaluations and paper experiment manifests.
 
 | Experiment | Entry |
 |---|---|
-| Sim2Real calibration | [30-case held-out calibration](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/7cf0da6324823aa98c8b4683c80f5dc75da0bb4f/experiments/sim2real-calibration) |
 | Visual history | [Visual history](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/visual-history) |
 | Latency in prompt | [Latency in prompt](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/latency-in-prompt) |
 | Latency transfer | [Latency transfer](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/main/experiments/latency-transfer) |

experiments/task-transfer/README.md

@@ -8,3 +8,5 @@
 Shared checkpoints and datasets have one canonical location in benchmark-models and benchmark-datasets. Fixed references and source limits are in each manifest. Latency profiles use the existing profiles repository.
 
 Verification uses fixed Git/LFS object identities and sizes. All 46 Parquet files passed real sample reads; five representative images decoded. Complete framework loading was not repeated.
+
+[Consolidation and source-retirement report](../../reports/task-transfer-20261007/README.md).

reports/task-transfer-20261007/README.md

@@ -0,0 +1,51 @@
+# Task transfer 资产合并
+
+正式收录 Demon Attack 零延迟 step 7000 和混合延迟 step 5000。
+Archive 收录 DefendTheLine + Deadly Corridor step 2000 和六任务混合训练 step 15000。
+
+正式入口：[task-transfer@e983a3a7](https://huggingface.co/datasets/latency-sensitive-bench/benchmark-datasets/tree/e983a3a72e042c6fab5575b847cdf6f4432cddaa/experiments/task-transfer)。
+共享模型和数据固定在 benchmark-models、benchmark-datasets，profiles 沿用现有仓库。
+
+## 收录与核验
+
+- 252 个源文件的目标 Git/LFS 身份和大小均匹配，包含权重、原始配置、训练状态、日志、数据和全部评测版本。
+- 46 个 Parquet 均通过 schema、行数和真实样本读取；五个代表图像样本解码通过。没有逐行读取全部数据。
+- 24 个正式评测单元均从逐局回报重算，每单元 20 局、seed 42–61。23 个点与论文整数显示一致。
+- 15 个具名数据选择项保留完整路径和实测 feature/schema、行数。128 个中央固定引用可解析。
+- 31 项相关测试通过，27 个 shell、7 段内嵌 Python、2 个 YAML 和 2 个 Python 文件检查通过。
+- 12 个消费者的旧 revision 与中央路径不匹配，已改为身份核验后的固定目标。
+- 两项既有 converter 断言失败与之前的 baseline 相同。没有为通过无关断言修改 converter。
+
+## 方法与限制
+
+源文件约 87.413 GB，其中约 87.395 GB 的 LFS 文件由 HF 服务端复制。
+纠正执行方式前，196 个普通 Git 文件、约 18.255 MB 内容经过了本地下载和 SDK 上传。
+纠正后，仅发送 Git 树、提交和新编写的说明元数据，复用 HF 已有对象。
+纠正后的 3 次成功元数据提交，共发送约 91.186 KB 压缩 pack。
+没有再次搬运已有权重或轨迹。
+四个固定 Deadly 延迟入口的精确数据选择已恢复，并加入 2/4/6/8 的回归用例。
+
+完整 framework 模型加载未重复验证。远端对象身份核验只证明源权重字节完整保留。
+正式分类按用户决定记录，不替代权重与论文评测的精确身份绑定。
+
+论文 CSV 的 mixed Demon Attack L0 为 5812，公开评测重算为 5832.25。
+论文称延迟为 raw frames，corrected mixed 配置中 label 4 为 266.67 ms，
+环境 60 FPS、观察 15 FPS，等于 16 raw frames。两项差异保留，论文数值未改。
+Mixed 原始训练配置、归一化统计和精确训练数据版本仍缺失。
+Archive 原配置没有固定训练数据 revision；available snapshot 不冒充精确训练字节绑定。
+
+## 源仓库清理
+
+已整仓删除 8 个组织源仓库：2 个 models、6 个 datasets。
+组织 API 统计由 models 47 / datasets 48 变为 models 45 / datasets 42。
+memory-rollouts 的其他实验资产继续保留，个人上游仓库保留。
+八个删除源均返回 404；删除后的 128 个固定引用仍可解析，八份正式评测摘要回读通过。
+没有重新训练或评测。完整 framework 加载保持未重复验证状态。
+
+主要回执：copy_manifest.json、current_asset_identity_verification.json、
+final_validation.json、paper_numeric_comparison.json、local_consumer_verification.json
+和 git_metadata_publication_*.json。
+
+论文来源说明已同步 Overleaf：`c829254`。
+
+一次性迁移工具和未完成的大文件下载将在回执发布后删除。

reports/task-transfer-20261007/final_validation.json

@@ -0,0 +1,19 @@
+{
+  "ready_for_retirement": true,
+  "source_and_target_object_identity_verified": 252,
+  "fixed_references": 128,
+  "published_metadata_documents": 74,
+  "named_dataset_selections": 15,
+  "parquet_samples": 46,
+  "decoded_images": 5,
+  "paper_comparison_cells": 24,
+  "paper_values_unchanged": true,
+  "root_navigation": {
+    "benchmark-models": true,
+    "benchmark-datasets": true,
+    "profiles": true
+  },
+  "local_consumers_verified": true,
+  "complete_framework_loading": "not repeated",
+  "verification_basis": "Remote Git/LFS object identity preserves the source checkpoint bytes; original inference metadata gaps remain explicit."
+}

reports/task-transfer-20261007/metadata_publication_history.json

@@ -0,0 +1,23 @@
+[
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "6266ddcd0d1764c112ee3dfb3836ef04bc860f01",
+    "compressed_pack_bytes": 65780,
+    "existing_artifact_payload_downloaded": false,
+    "existing_artifact_payload_uploaded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "9bd2fe49938f144d4c06325af0406819ab5b5da8",
+    "compressed_pack_bytes": 10341,
+    "existing_artifact_payload_downloaded": false,
+    "existing_artifact_payload_uploaded": false
+  },
+  {
+    "repo": "latency-sensitive-bench/benchmark-datasets",
+    "revision": "e983a3a72e042c6fab5575b847cdf6f4432cddaa",
+    "compressed_pack_bytes": 15065,
+    "existing_artifact_payload_downloaded": false,
+    "existing_artifact_payload_uploaded": false
+  }
+]

reports/task-transfer-20261007/occam_review.json

@@ -0,0 +1,35 @@
+{
+  "review_method": "fresh-context read-only review of the task-owned consumer diff",
+  "resolved_findings": [
+    {
+      "classification": "fix",
+      "behavior": "Preserve each fixed Deadly L2/L4/L6/L8 training data selection.",
+      "evidence": "exact_dataset_selection_repairs.json"
+    },
+    {
+      "classification": "delete",
+      "behavior": "Remove unused HF_REPO_ID from the six-task launcher."
+    },
+    {
+      "classification": "delete",
+      "behavior": "Remove retired per-task repository fallback from the rollout publisher."
+    }
+  ],
+  "new_test": {
+    "classification": "keep",
+    "requirement": "Regression: changing repository locations must preserve fixed-latency training dataset selection.",
+    "cases": [
+      2,
+      4,
+      6,
+      8
+    ],
+    "owner": "StarVLA rl_games training commands",
+    "lifecycle": "long-lived contract"
+  },
+  "unresolved_findings": [],
+  "residual_limits": [
+    "Complete framework loading was not repeated.",
+    "Historical training data revisions remain unconfirmed when the original run did not pin them."
+  ]
+}

reports/task-transfer-20261007/post_retirement_verification.json

@@ -0,0 +1,50 @@
+{
+  "retired_sources": [
+    {
+      "repo": "latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/starvla-demon-attack-zero-latency-evals",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/starvla-demon-attack-mixed-024-latency-evals",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/deadly_1000ep",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/asterix_200ep",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/atlantis_200ep",
+      "status": 404
+    },
+    {
+      "repo": "latency-sensitive-bench/defend_the_line_200ep",
+      "status": 404
+    }
+  ],
+  "models_removed": 2,
+  "datasets_removed": 6,
+  "organization_before": {
+    "models": 47,
+    "datasets": 48
+  },
+  "organization_after": {
+    "models": 45,
+    "datasets": 42
+  },
+  "fixed_references_resolve": 128,
+  "actual_small_evaluation_readbacks": 8,
+  "source_checkpoint_identities_preserved": true,
+  "large_artifact_payloads_downloaded": false
+}

reports/task-transfer-20261007/repository_deletions.json

@@ -0,0 +1,58 @@
+[
+  {
+    "repo": "latency-sensitive-bench/openvla_bridge_cross_defendtheline_zero_deadly_zero_exp3",
+    "repo_type": "model",
+    "source_revision": "c4a60d71edd7ecd15af5b63b764f8c0742e0d3f7",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/openvla_bridge_cross_mixed_all_024_5env200ep_deadly1000ep_exp1",
+    "repo_type": "model",
+    "source_revision": "53eb98e6177d630b3ee9775c84413300d7b0f29b",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/starvla-demon-attack-zero-latency-evals",
+    "repo_type": "dataset",
+    "source_revision": "c85322cb658371a46195d4acd34337373a2f7d28",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/starvla-demon-attack-mixed-024-latency-evals",
+    "repo_type": "dataset",
+    "source_revision": "602487c53e6aec515e61aebb7afb44845f6ad468",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/deadly_1000ep",
+    "repo_type": "dataset",
+    "source_revision": "fa0c35ef651bb1f938d3830f18922a0383c3fb13",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/asterix_200ep",
+    "repo_type": "dataset",
+    "source_revision": "726b2fc79ff0509d8819e564ab2695daa5215ced",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/atlantis_200ep",
+    "repo_type": "dataset",
+    "source_revision": "447f7bf629302e9646c24c7e97a03b3a5d5ad8d2",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  },
+  {
+    "repo": "latency-sensitive-bench/defend_the_line_200ep",
+    "repo_type": "dataset",
+    "source_revision": "1f393d6e99cac2ec9d4fedd009d520bb6695ab6a",
+    "deleted": true,
+    "target_verification": "final_validation.json"
+  }
+]

reports/task-transfer-20261007/test_verification.json

@@ -0,0 +1,22 @@
+{
+  "training_config_tests": 20,
+  "checkpoint_and_evaluation_tests": 11,
+  "known_existing_converter_failures": [
+    "test_convert_flappy_hf_loader_passes_dataset_source_subdir",
+    "test_convert_flappy_index_split_retries_canonical_hf_columns"
+  ],
+  "paper_arxiv_build": "up-to-date; no numeric/figure changes",
+  "new_regression_test": {
+    "file": "third_party/starVLA/tests/rl_games/test_training_commands.py",
+    "name": "test_deadly_fixed_training_preserves_latency_dataset",
+    "cases": [
+      2,
+      4,
+      6,
+      8
+    ],
+    "contract": "Fixed-delay launchers retain their exact pre-migration training dataset selection.",
+    "lifecycle": "long-lived contract",
+    "owner": "StarVLA rl_games training command suite"
+  }
+}