Release validated Zebra seed0 and simplify standard HF workflow

README.md

@@ -49,18 +49,19 @@ corresponding model folder from this repository.
 |---|---:|---|---:|---:|
 | sudoku | 0 | EqR Sudoku-Lite, 2,048 puzzles | 99.7070% (2042/2048) | 899.75 |
 | maze_unique | 0 | Official EqR Maze-Unique, 1,000 puzzles | 99.9% (999/1000) | 15.708 |
-| zebra | 1 | 5x5 Zebra, 1,024-puzzle seed0 subset | 99.4141% (1018/1024) | 4.0879 |
+| zebra | 0 | 5x5 Zebra, 1,024-puzzle seed0 subset | 99.6094% (1020/1024) | 6.2402 |
 
 Sudoku costs 0.8062 mean TFLOPs/puzzle; Maze-Unique costs 0.1781. Zebra FLOPs
-are unprofiled for this preprocessing, so only NFE is reported. Earlier Zebra SDE
-inference scored 99.8047% (1022/1024) on the same checkpoint; its 257-NFE held-t
-trajectory scores 99.7070%. The early-stop/restart policy trades some accuracy for
-substantially less compute. All Zebra test puzzles accept on the first attempt.
+are unprofiled for this preprocessing, so only NFE is reported. Zebra seed0 scores99.8047% at257NFE with the fixed-depth held-t trajectory;
+its stopping/restart policy scores99.6094% at6.2402meanNFE. Its earlier SDE
+result is99.6094%. Every test puzzle accepts on the first attempt. FLOPs for this
+5x5 setup are unprofiled; NFE is reported instead.
 
-These are selected-seed results, not seed averages. Maze-Unique restart seeds0/1/2:
+All three hosted checkpoints now use training seed0. Maze-Unique restart seeds0/1/2:
 99.9%,98.5%,99.8%. Zebra SDE seeds0/1/2:99.6094%,99.8047%,99.5117%.
-Only one new Sudoku training seed was run. The Zebra checkpoint is selected using
-that recorded SDE evaluation, not a comparison of restart policies across seeds.
+Only one new Sudoku training seed was run. Zebra seed0 was preferred for a
+consistent release after its measured default score99.6094% exceeded seed1's
+99.4141%. Source checkpoint/step/seed details remain in each manifest.
 
 ## Scope and reproducibility
 

maze_unique/manifest.json

@@ -19,5 +19,5 @@
     "calibration.npz": "3f0d8261ef2977c8a8d219dcb3f6cd146d3e7195438cfbe8939e774d1d8ff0dc",
     "source_evaluation.json": "cc68b4b64c9aa49b7a829ba36206dc05fd928668a349955b9263c3a1a4b4bcac"
   },
-  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
+  "code_revision": "e41dd277d4176fbc29b39219a1aef5c60a5481d8"
 }

release.json

@@ -1,8 +1,13 @@
 {
   "code_repo": "helblazer811/FlowMapReasoning",
   "code_ref": "public-simplified",
-  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3",
+  "code_revision": "e41dd277d4176fbc29b39219a1aef5c60a5481d8",
   "package": "frm",
   "pypi_published": false,
-  "scope": "private draft"
+  "scope": "private draft",
+  "training_seeds": {
+    "sudoku": 0,
+    "maze_unique": 0,
+    "zebra": 0
+  }
 }

sudoku/manifest.json

@@ -19,5 +19,5 @@
     "calibration.npz": "e1f3b99cd5ce91dfe1e7de44995e1e38bbf1372766457025eca589625e89281e",
     "source_evaluation.json": "6085f2d71fe4aa39d253820326922de1236a142527fabf5746c8202269e0251c"
   },
-  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
+  "code_revision": "e41dd277d4176fbc29b39219a1aef5c60a5481d8"
 }

verification.json

@@ -1,24 +1,41 @@
 {
-  "mode": "functional smoke, not full benchmark re-evaluation",
-  "device": "cpu",
-  "num_puzzles_per_task": 1,
-  "installed_from": "GitHub/public-simplified",
-  "package": "frm",
-  "package_version": "0.2.0",
-  "tested_code_revision": "5c40fd89631c5f7384589b408f6591dd98057621",
-  "tested_bundle_revision": "3e8c45fd28b56f65c072334fc25c51328aa57850",
-  "tasks": {
-    "maze_unique": {
-      "passed": true,
-      "model_sha256": "3f0023f50352d395ca97a1d4a5c08a0f4bae4140fb53411a0a9c3840f7a7a4e3"
-    },
-    "sudoku": {
-      "passed": true,
-      "model_sha256": "5952ab8fb4184cffa7580b2035bc269531350d27e95cbdb4d26c9d5b241066c2"
-    },
-    "zebra": {
-      "passed": true,
-      "model_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4"
+  "previous_download_smoke": {
+    "mode": "functional smoke, not full benchmark re-evaluation",
+    "device": "cpu",
+    "num_puzzles_per_task": 1,
+    "installed_from": "GitHub/public-simplified",
+    "package": "frm",
+    "package_version": "0.2.0",
+    "tested_code_revision": "5c40fd89631c5f7384589b408f6591dd98057621",
+    "tested_bundle_revision": "3e8c45fd28b56f65c072334fc25c51328aa57850",
+    "tasks": {
+      "maze_unique": {
+        "passed": true,
+        "model_sha256": "3f0023f50352d395ca97a1d4a5c08a0f4bae4140fb53411a0a9c3840f7a7a4e3"
+      },
+      "sudoku": {
+        "passed": true,
+        "model_sha256": "5952ab8fb4184cffa7580b2035bc269531350d27e95cbdb4d26c9d5b241066c2"
+      },
+      "zebra": {
+        "passed": true,
+        "model_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4"
+      }
     }
+  },
+  "standard_download_workflow": {
+    "code_revision": "700c1df787b00768bd3a3ef456a41a608ab2b343",
+    "method": "huggingface_hub.snapshot_download(local_dir) then python -m frm.evaluate",
+    "device": "cpu",
+    "task": "sudoku",
+    "num_puzzles": 1,
+    "passed": true
+  },
+  "zebra_seed0_gpu_validation": {
+    "job": 404162,
+    "num_puzzles": 1024,
+    "solved": 1020,
+    "mean_nfe": 6.240234375,
+    "model_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57"
   }
 }

zebra/expected_results.json

@@ -1,9 +1,9 @@
 {
   "schema_version": 1,
   "git_commit": null,
-  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/hub_release_20261018_v2/bundles/zebra/model.pt",
-  "checkpoint_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
-  "benchmark": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/hub_release_20261018_v2/bundles/zebra/test.npz",
+  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/model.pt",
+  "checkpoint_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
+  "benchmark": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/test.npz",
   "benchmark_sha256": "0948626c109bafdff92f16c8ea3a2c8a6078c39960c1261fed28d3a7f7204431",
   "policy_config": {
     "max_nfe": 257,
@@ -24,16 +24,16 @@
   "residual": "trailing-window mean of per-puzzle euclidean distance between consecutive denoiser distributions, averaged over non-clue cells",
   "oracle_used_for_test_selection": false,
   "kind": "residual_guided_inference_evaluation",
-  "run_fingerprint": "950d7c7190c686a8b3c3269f58ec1a1e936f91db6448454553f81bd0485ee9ca",
+  "run_fingerprint": "ad3a2420cab1679dab9f04d8f963fc6acf0bfce272b1c233e0e0d84b10810112",
   "num_puzzles": 1024,
   "max_samples_override": null,
-  "model_step": 20000,
-  "calibration": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/hub_release_20261018_v2/bundles/zebra/threshold.json",
-  "calibration_sha256": "36c26f2a36371bd2d0ad0d3924a501c778cf6ee8112d8b3138770835440ec8b7",
+  "model_step": 115000,
+  "calibration": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/threshold.json",
+  "calibration_sha256": "d7401816ce4e2638c587f5793f55b60a1110acb03a87941f71b1b05aa7bb46c3",
   "calibration_benchmark_sha256": "b2c66c97d53152de289a2ab0c4ad2123ec5bed4bf53aaab14f21fa7f90c527cd",
-  "threshold": 0.20256999680399107,
+  "threshold": 0.10050372427701933,
   "threshold_selection": {
-    "threshold": 0.20256999680399107,
+    "threshold": 0.10050372427701933,
     "accuracy": 1.0,
     "solved": 256,
     "mean_nfe": 4.51171875,
@@ -45,26 +45,26 @@
   "policies": [
     {
       "policy": "Fixed-depth FRM",
-      "solved": 1021,
+      "solved": 1022,
       "num_puzzles": 1024,
-      "solve_rate": 0.9970703125,
+      "solve_rate": 0.998046875,
       "mean_nfe": 257.0,
       "mean_tflops": null
     },
     {
       "policy": "+ Residual early stopping",
-      "solved": 1018,
+      "solved": 1020,
       "num_puzzles": 1024,
-      "solve_rate": 0.994140625,
-      "mean_nfe": 4.087890625,
+      "solve_rate": 0.99609375,
+      "mean_nfe": 6.240234375,
       "mean_tflops": null
     },
     {
       "policy": "+ Sequential restarts",
-      "solved": 1018,
+      "solved": 1020,
       "num_puzzles": 1024,
-      "solve_rate": 0.994140625,
-      "mean_nfe": 4.087890625,
+      "solve_rate": 0.99609375,
+      "mean_nfe": 6.240234375,
       "mean_tflops": null
     }
   ],
@@ -76,22 +76,21 @@
       "attempted": 1024,
       "accepted": 1024,
       "remaining": 0,
-      "mean_nfe_among_attempted": 4.087890625
+      "mean_nfe_among_attempted": 6.240234375
     }
   ],
   "restart_frontier": [
     {
       "breadth_cap": 1,
-      "solved": 1018,
+      "solved": 1020,
       "num_puzzles": 1024,
-      "solve_rate": 0.994140625,
-      "mean_nfe": 4.087890625,
+      "solve_rate": 0.99609375,
+      "mean_nfe": 6.240234375,
       "mean_tflops": null,
       "accepted": 1024,
       "unresolved": 0
     }
   ],
   "selection_rule": "stop each trajectory once the trailing-mean residual remains <= threshold for 1 checks and the decoded grid remains unchanged; restart only trajectories that reach max_nfe without acceptance; if max_breadth is exhausted, return the attempted terminal prediction with minimum residual",
-  "labels_used_for": "final scoring only; never stopping, restart, or candidate selection",
-  "compute_note": "FLOPs for this 5x5 model have not been profiled; use measured NFE."
+  "labels_used_for": "final scoring only; never stopping, restart, or candidate selection"
 }

zebra/manifest.json

@@ -1,25 +1,25 @@
 {
   "task": "zebra",
-  "training_seed": 1,
-  "selection": "best measured seed among recorded runs; see results table",
-  "source_checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s1/stage_b/best.pt",
-  "source_checkpoint_sha256": "e1e8c364539adf482d04d0de4e234e7a1bfc71e55547824d2dd5789f85fa0ef4",
-  "source_step": 20000,
-  "checkpoint_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
+  "training_seed": 0,
+  "selection": "seed0 release for consistency; see complete seed results",
+  "source_checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s0/stage_b/best.pt",
+  "source_checkpoint_sha256": "c2c566d99eff64788bfbbfb8dc4571a6f0b451f136fffd772298e56ac11c80cd",
+  "source_step": 115000,
+  "checkpoint_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
   "checkpoint_format": "raw model plus EMA/config/provenance; optimizer omitted",
   "scope": "5x5 Zebra, repacked answers (not paper mixed-size pinned)",
+  "calibration_note": "256 unique 5x5 training rows, seed20260910, excludes test puzzles",
   "files": {
-    "training.yaml": "29d7da5265a805190959d234de1f3702591c15632b80ac713f3744706147d3c2",
-    "model.pt": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
-    "provenance.json": "476fdce0c4bdd67e05c3e7554c9a6f99a9b54406eac859cc59ca48dc293e8974",
-    "expected_results.json": "46a370d09021d2f900475851296b17fcd7fbe332dd5b4ac48fe7878dcf051cfd",
-    "test.npz": "0948626c109bafdff92f16c8ea3a2c8a6078c39960c1261fed28d3a7f7204431",
-    "threshold.json": "36c26f2a36371bd2d0ad0d3924a501c778cf6ee8112d8b3138770835440ec8b7",
-    "vocab.json": "17c8b60b8a2ec9feba73a0d6f5cd5f973da60c54d0d82b23a61743c27715698d",
     "calibration.npz": "b2c66c97d53152de289a2ab0c4ad2123ec5bed4bf53aaab14f21fa7f90c527cd",
+    "vocab.json": "17c8b60b8a2ec9feba73a0d6f5cd5f973da60c54d0d82b23a61743c27715698d",
     "meta.json": "1d19448cf8bc66d7bb87d8d5ac73367cc10c61eac77acf5172ba1c5dcd10130c",
-    "source_evaluation.json": "8466835dc4e90919c24cfefcc5de98814544e14c42b9f076631fda0d97397375"
+    "model.pt": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
+    "expected_results.json": "80a00058c2e5edb8ed4bacb9d469bcdf41050554266c80d36e4600e575d8ece9",
+    "provenance.json": "0895404bd22920def2dcccdf0d3fc0af3e11efaa4d65c285a3f68e3def6a6b9b",
+    "test.npz": "0948626c109bafdff92f16c8ea3a2c8a6078c39960c1261fed28d3a7f7204431",
+    "training.yaml": "83826d69f8f110cb4b5b6f4477c2a1578524c2326d05b4fd160f8c733dd6f2db",
+    "threshold.json": "d7401816ce4e2638c587f5793f55b60a1110acb03a87941f71b1b05aa7bb46c3",
+    "source_evaluation.json": "68e5262551fd44859df69376119baf15e68ed5eb368a34e96d61f6fcbb17ffd9"
   },
-  "calibration_note": "256 unique 5x5 training rows, seed20260910, explicitly excludes test puzzles; not the mixed-size paper policy.",
-  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
+  "code_revision": "e41dd277d4176fbc29b39219a1aef5c60a5481d8"
 }

zebra/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4
-size 200110615
+oid sha256:c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57
+size 200110743

zebra/provenance.json

@@ -1,11 +1,11 @@
 {
   "task": "zebra",
-  "training_seed": 1,
-  "selection": "best measured seed among recorded runs; see results table",
-  "source_checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s1/stage_b/best.pt",
-  "source_checkpoint_sha256": "e1e8c364539adf482d04d0de4e234e7a1bfc71e55547824d2dd5789f85fa0ef4",
-  "source_step": 20000,
-  "checkpoint_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
+  "training_seed": 0,
+  "selection": "seed0 release for consistency; see complete seed results",
+  "source_checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s0/stage_b/best.pt",
+  "source_checkpoint_sha256": "c2c566d99eff64788bfbbfb8dc4571a6f0b451f136fffd772298e56ac11c80cd",
+  "source_step": 115000,
+  "checkpoint_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
   "checkpoint_format": "raw model plus EMA/config/provenance; optimizer omitted",
   "scope": "5x5 Zebra, repacked answers (not paper mixed-size pinned)"
 }

zebra/source_evaluation.json

@@ -1,7 +1,7 @@
 {
   "dataset": "zebra",
-  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s1/stage_b/best.pt",
-  "checkpoint_step": 20000,
+  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/20261015-1119/full/zebra_s0/stage_b/best.pt",
+  "checkpoint_step": 115000,
   "use_ema": true,
   "protocol": {
     "split": "test",
@@ -19,15 +19,15 @@
   },
   "metrics": {
     "n_samples": 1024,
-    "exact_match": 0.998046875,
-    "solved": 0.998046875,
-    "token_accuracy": 0.999140625,
+    "exact_match": 0.99609375,
+    "solved": 0.99609375,
+    "token_accuracy": 0.9985546875,
     "clue_consistency": 1.0
   },
   "checkpoint_provenance": {
     "dataset": "zebra",
     "stage": "b",
-    "seed": 1,
+    "seed": 0,
     "recipe": {
       "data": {
         "dataset": "zebra",
@@ -120,7 +120,7 @@
         "self_cond": true
       },
       "pipeline": {
-        "seed": 1,
+        "seed": 0,
         "num_workers": 2,
         "stage_a_steps": 200000,
         "stage_b_steps": 200000,

zebra/threshold.json

@@ -2,7 +2,7 @@
   "schema_version": 1,
   "git_commit": null,
   "checkpoint": "model.pt",
-  "checkpoint_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
+  "checkpoint_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
   "benchmark": "calibration.npz",
   "benchmark_sha256": "b2c66c97d53152de289a2ab0c4ad2123ec5bed4bf53aaab14f21fa7f90c527cd",
   "policy_config": {
@@ -26,10 +26,10 @@
   "kind": "residual_stopping_calibration",
   "num_puzzles": 256,
   "max_samples_override": null,
-  "model_step": 20000,
+  "model_step": 115000,
   "labels_used_for": "development-set threshold calibration only",
   "selected": {
-    "threshold": 0.20256999680399107,
+    "threshold": 0.10050372427701933,
     "accuracy": 1.0,
     "solved": 256,
     "mean_nfe": 4.51171875,

zebra/training.yaml

@@ -22,7 +22,7 @@ train:
   val_every: 5000
   save_best_metric: val/solved
 pipeline:
-  seed: 1
+  seed: 0
   num_workers: 2
   stage_a_steps: 200000
   stage_b_steps: 200000