Publish verified formal StatePlay model and reproducibility metadata

PUBLICATION.json

@@ -0,0 +1,32 @@
+{
+  "schema": "stateplay-formal-model-publication-v1",
+  "repo": "lupele/rl_model",
+  "created_at": "2026-10-22T01:22:28.504083+00:00",
+  "kind": "formal_stage_A",
+  "run_id": "study_v1",
+  "conditions": 2280,
+  "candidates": 9120,
+  "optimizer_step": 652,
+  "mode": "B-only",
+  "smoke": false,
+  "source_sha256": "0de4a219b2e81424a909734648e55da602106586eb9638aa58e4522ed51b57bd",
+  "published_sha256": "56328f45f3ee19237b63add2eb666dfcaa8aaebf505d29aa400a78519aea28ac",
+  "metadata_only_changes": [
+    "/config/reward/endpoint",
+    "/config/reward_model/inference_url",
+    "/config/config_hash"
+  ],
+  "preserved_exactly": [
+    "schema",
+    "live",
+    "policies",
+    "optimizer",
+    "scheduler",
+    "scaler",
+    "state",
+    "rank_rng",
+    "sft_sha256",
+    "tokenizer_sha256"
+  ],
+  "all_parameters_optimizer_rng_exactly_equal": true
+}

README.md

@@ -1,3 +1,29 @@
 ---
-license: apache-2.0
+base_model: lupele/ourmodel
+tags: [stateplay, diffusionnft, reinforcement-learning, world-model]
 ---
+# StatePlay 正式 Stage A DiffusionNFT checkpoint
+
+`study_v1` 正式 Stage A 最终状态：**2,280 条件、9,120 候选视频、652 次优化器更新、B-only**。
+这是正式训练结果；不是 smoke 或四卡调试模型，不代表 Stage C 已完成。
+
+`world.pt` 是增量训练状态，包含 live/old/ref、optimizer、RNG、训练进度。
+它只更新 latent state transition，不是完整 DiT；推理先加载配套 SFT 和 tokenizer，
+再叠加 `world.pt['live']`。续训还须恢复优化器等状态，并从数据仓库恢复条件顺序和消费账本。
+`world.json` 是公开文件的正确校验 sidecar。
+
+公开副本仅将内嵌 API/本地评分服务地址替换为 `.invalid` 占位地址并重算配置哈希。
+所有参数、live/old/ref、优化器、RNG、652 步进度均逐项验证完全一致。
+重新序列化使文件 SHA256 与历史原件不同；两者映射见 `PUBLICATION.json`。
+
+旧 A100 转移包固定了原件 SHA256，直接传入公开副本会触发身份校验失败。
+请在转移包**复制件**中依据 `PUBLICATION.json` 更新这项身份及包完整性索引；保留校验。
+迁移还须映射本机绝对路径，保留算法配置和账本。API 地址与密钥由本机环境单独提供。
+
+
+代码：[lupele/stateplay-test](https://github.com/lupele/stateplay-test) 的 `rl_posttrain/`。
+数据：[StatePlay-Reward-2280](https://huggingface.co/datasets/lupele/StatePlay-Reward-2280)。
+配套模型：[SFT](https://huggingface.co/lupele/ourmodel)、[transition tokenizer](https://huggingface.co/lupele/transition_tokenizer.pt)、[Stage A 更新](https://huggingface.co/lupele/rl_model)、[Reward Model](https://huggingface.co/lupele/rewardmodel)。
+
+身份与发布处理见 `PUBLICATION.json`，公开文件 SHA256 见 `SHA256SUMS`。
+原始本地文件保持原位且未修改。仓库不包含 API 凭据或真实服务地址。

SHA256SUMS

@@ -0,0 +1,4 @@
+775fe06d2ccd691dd527cb06f6902a14a89621ed3a03441767fc5cfceea4e367  PUBLICATION.json
+deb0af344c8011adc5b930c89a3bd15889aed6f2701cd068f54d8eecd7fc0a59  README.md
+827497a4c072985ad9662bf1f1e6e2ba52a0823b3f7f2a652d8dc67c373b575f  world.json
+56328f45f3ee19237b63add2eb666dfcaa8aaebf505d29aa400a78519aea28ac  world.pt

world.json

@@ -0,0 +1,39 @@
+{
+  "sha256": "56328f45f3ee19237b63add2eb666dfcaa8aaebf505d29aa400a78519aea28ac",
+  "schema": "rl-world-checkpoint-v1",
+  "state": {
+    "optimizer_step": 652,
+    "batch_id": 81,
+    "update_cursor": 0,
+    "batch_done": true,
+    "stage": "api",
+    "smoke": false,
+    "backend": {
+      "deterministic_algorithms": true,
+      "cudnn_deterministic": true,
+      "cudnn_benchmark": false,
+      "cublas_workspace_config": ":4096:8",
+      "gradient_reduction": "name-ordered-fp32-global-mean-v2",
+      "torch_version": "2.4.1+cu124"
+    },
+    "schedule_world_size": 7,
+    "ema_eta": 0.5,
+    "all_filtered_batch": false,
+    "rng_topology_mapping": {
+      "saved_visible_devices": 7,
+      "current_visible_devices": 7,
+      "kept_logical_devices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6
+      ]
+    }
+  },
+  "mode": "B-only",
+  "config_hash": "74ca72fa2f0970e3d7364ac36593e02514d9585512294cbf92c876e7337bad4a",
+  "sft_sha256": "53eff9e5e59e92bb3cc7810e3ff36f9ee490be4a122a3a8ff96a4de0a03b4c5d"
+}

world.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56328f45f3ee19237b63add2eb666dfcaa8aaebf505d29aa400a78519aea28ac
+size 33458850