{
  "schema_version": "edgeflowertune.reference_assets.v1",
  "status": "recovered_reference_fingerprints_not_official_release",
  "date": "2026-10-19",
  "basis": "Recovered 2026-10-18 reference-rerun files, not proof that every legacy leaderboard score used identical bytes.",
  "models": {
    "qwen25_05b": {
      "repository": "Qwen/Qwen2.5-0.5B",
      "variant": "pretrained/base",
      "revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
      "revision_basis": "All seven recorded weight/config/tokenizer file hashes match the independently downloaded 2026-10-19 snapshot. This identifies matching content, not the historical download date.",
      "files": {
        "config.json": "479dcf0c5286339e41ad3992cd08ae88a467c4187587936248e2b7c96283484b",
        "tokenizer.json": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539",
        "model.safetensors": "88c142557820ccad55bb59756bfcfcf891de9cc6202816bd346445188a0ed342",
        "generation_config.json": "8c970692323e3ea0e9b8b0a4dca79388d31226e41f83c9fd6014804280ebf6e8",
        "vocab.json": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910",
        "merges.txt": "599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3",
        "tokenizer_config.json": "c91efca15ceff6e9ee9424db58a6f59cd41294e550a86cbd07e3c1fb500b34f9"
      }
    },
    "gemma3_270m": {
      "repository": "google/gemma-3-270m-it",
      "variant": "instruction-tuned",
      "revision": "ac82b4e820549b854eebf28ce6dedaf9fdfa17b3",
      "revision_basis": "Downloaded at the locally archived HF revision",
      "files": {
        "config.json": "8983343e5bfbf8e336d222a522a1d781443491e9de2d081b8856f8fd26ab6046",
        "tokenizer.json": "7d4046bf0505a327dd5a0abbb427ecd4fc82f99c2ceaa170bc61ecde12809b0c",
        "model.safetensors": "700b710a9a99c295ed546647aa81cacf9f81f4c573ea2be613a0e2517a44afab",
        "added_tokens.json": "50b2f405ba56a26d4913fd772089992252d7f942123cc0a034d96424221ba946",
        "chat_template.jinja": "7de1c58e208eda46e9c7f86397df37ec49883aeece39fb961e0a6b24088dd3c4",
        "tokenizer_config.json": "94b03056ec5831e9021c3e3fe9db682778c4f2d081c4188f1fcb8b90541cf1cf",
        "tokenizer.model": "1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c",
        "special_tokens_map.json": "2f7b0adf4fb469770bb1490e3e35df87b1dc578246c5e7e6fc76ecf33213a397",
        "generation_config.json": "f3d7bb125d6b8751b8fc57c7155c10d6716ebe5a64900a27f56ca236c1c0c241"
      }
    },
    "gemma3_1b": {
      "repository": "google/gemma-3-1b-pt",
      "variant": "pretrained/base",
      "revision": "fcf18a2a879aab110ca39f8bffbccd5d49d8eb29",
      "revision_basis": "On 2026-10-19 all eight archived file hashes matched both the rehashed preserved GPU-server copy and an independently downloaded full snapshot of this pinned official revision. The participant --reference command and offline verification passed. This identifies matching content, not the historical download commit or every legacy score.",
      "files": {
        "config.json": "4dd5218cebc4fdea9004b944d34672a2d47f9ef930e3bb16a3fc4bbbe0fcee81",
        "tokenizer.json": "7d4046bf0505a327dd5a0abbb427ecd4fc82f99c2ceaa170bc61ecde12809b0c",
        "model.safetensors": "ee5250f6eb1aa7cfb729dfd4dc8d9964fd772324776c6d00bf2bc674c069cb27",
        "added_tokens.json": "50b2f405ba56a26d4913fd772089992252d7f942123cc0a034d96424221ba946",
        "generation_config.json": "fd9324becc53c4be610db39e13a613006f09fd6ef71a95fb6320dc33157490a3",
        "tokenizer.model": "1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c",
        "special_tokens_map.json": "2f7b0adf4fb469770bb1490e3e35df87b1dc578246c5e7e6fc76ecf33213a397",
        "tokenizer_config.json": "4d97b6d876fd05ffaf72c3243231975c50263ad4fc620bc9f0b44be2ca11fa31"
      }
    }
  },
  "datasets": {
    "boolq": {
      "train_rows": 9427,
      "validation_rows": 3270,
      "input_sha256": {
        "train": "909d15b34f51045e5fa9f51c256f2af092b0378878ffa25688d0ccbce7f432e7",
        "validation": "13e1aad55f89fd431782902360c2ea6d7dcb31587050358c39aa7baa8ed230e1"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": "3de24cf8022e94f4ee4b9d55a6f539891524d646",
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "huggingface_parquet",
        "repository": "aps/super_glue",
        "original_repository": "super_glue",
        "configuration": "boolq",
        "revision": "3de24cf8022e94f4ee4b9d55a6f539891524d646",
        "archived_metadata_sha256": "09ae866c293a476d6631ae579c7612e39e2a22e31867f7319a3bf49d8e30dc4d",
        "files": {
          "train": {
            "path": "boolq/train-00000-of-00001.parquet",
            "sha256": "41e2ce07ac0d3bcd411bd4b6326bab2cff5fa67bacdbbcea74bb9e001ec635d3",
            "bytes": 3846588
          },
          "validation": {
            "path": "boolq/validation-00000-of-00001.parquet",
            "sha256": "f46438b8cce048184552a2bfcdb8252eadb3a2602f0c22334caccd9b6ea547d6",
            "bytes": 1313731
          }
        }
      },
      "exporter_key": "boolq",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "efa9a103e563f74f358839c8d49463b9e43415a72301e6f1af36061fac46e221",
        "validation": "707dfa1910a8ab337472aa734ed58ac3cac00112f50b5d7847a8e2e6bc2283cc"
      },
      "unique_sample_ids": {
        "train": 9427,
        "validation": 3270
      }
    },
    "qnli": {
      "train_rows": 104743,
      "validation_rows": 5463,
      "input_sha256": {
        "train": "73c924a7009ccd9000e96894ac7778ec706f47307526204cf2b101b6634db92c",
        "validation": "3d317752b4963ad23f7aa819e28a3d7f2bb5ad52414c7b4ef707f5ad94351543"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": "bcdcba79d07bc864c1c254ccfcedcce55bcc9a8c",
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "huggingface_parquet",
        "repository": "nyu-mll/glue",
        "original_repository": "glue",
        "configuration": "qnli",
        "revision": "bcdcba79d07bc864c1c254ccfcedcce55bcc9a8c",
        "archived_metadata_sha256": "e88b73f0f378e15e4cd4c68f7ac7b3815a895061a40c373d3457b9d7b5429bc2",
        "files": {
          "train": {
            "path": "qnli/train-00000-of-00001.parquet",
            "sha256": "ebc7cb70a5bbde0b0336c3d51f31bb4df4673e908e8874b090b52169b1365c6c",
            "bytes": 17528917
          },
          "validation": {
            "path": "qnli/validation-00000-of-00001.parquet",
            "sha256": "e69311b81dc65589286091d9905a27617a90436dd215c7a59832fa8f4f336169",
            "bytes": 872062
          }
        }
      },
      "exporter_key": "qnli",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "ce141d14570105eaeaf7207dd17687ca98d48ef8f9dbd88503cb3e580a2c2fc4",
        "validation": "f514ecdf13b813259981e7ebe7c4dcabc6bfc9ecb2d380f088e5e7df997e9885"
      },
      "unique_sample_ids": {
        "train": 104743,
        "validation": 5463
      }
    },
    "piqa": {
      "train_rows": 16113,
      "validation_rows": 1838,
      "input_sha256": {
        "train": "c189d2c5d745d51879e72be582f5b98430e2a9cfe8746434cf9d4d8c963d581b",
        "validation": "4003fa363e65085b026927f5939d27d6e12c754409001d431b25f635a2f3a44c"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": null,
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "official_zip",
        "url": "https://storage.googleapis.com/ai2-mosaic/public/physicaliqa/physicaliqa-train-dev.zip",
        "sha256": "54d32a04f59a7e354396f321723c8d7ec35cc6b08506563d8d1ffcc15ce98ddd",
        "bytes": 1824009,
        "revision": null,
        "revision_basis": "Content-pinned archive: freshly fetched bytes match the historical ZIP; no Git revision applies."
      },
      "exporter_key": "piqa",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "895c0f99fbc15cb553c3e6c7b661dde10024f806d8332443d32c7ddd73f21e48",
        "validation": "91d3a72168e750e1f3e8c0117d6da92e73118cfe3577b6711d0e5c110a770543"
      },
      "unique_sample_ids": {
        "train": 16113,
        "validation": 1838
      }
    },
    "hellaswag": {
      "train_rows": 39905,
      "validation_rows": 10042,
      "input_sha256": {
        "train": "e16c7acf8ba76a4f7fc0a21420f242c5e0a69bd1c066fd64750a2dff66a5cc38",
        "validation": "60f4f66fe7b0a10c64376a69ecc179e1bcad0d72dc5756791348af735f10eff3"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": "218ec52e09a7e7462a5400043bb9a69a41d06b76",
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "huggingface_parquet",
        "repository": "Rowan/hellaswag",
        "original_repository": "Rowan/hellaswag",
        "configuration": "default",
        "revision": "218ec52e09a7e7462a5400043bb9a69a41d06b76",
        "archived_metadata_sha256": "ad8876907b93176bcda467bfe8cb9ee2c20b3271d1dc015a1675b61836f2c712",
        "files": {
          "train": {
            "path": "data/train-00000-of-00001.parquet",
            "sha256": "cacb12587faa63d7f723a72d61d12bfa94b140446f5a6a0a2e1c6906ab88bf02",
            "bytes": 24365524
          },
          "validation": {
            "path": "data/validation-00000-of-00001.parquet",
            "sha256": "899813071e1e95efafec90f856e1987d2150fa4d020fc005df6962c259f660cd",
            "bytes": 6315951
          }
        }
      },
      "exporter_key": "hellaswag",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "def39f0dc39b2c0da96249493b7764179bec4d96bd059eb85ef0140866b31ee4",
        "validation": "fc5a9573bdbb508b4ce9bc68e8093721fdab4ab1be8ae802d3ce5672a18aec71"
      },
      "unique_sample_ids": {
        "train": 32599,
        "validation": 9609
      }
    },
    "socialiqa": {
      "train_rows": 33410,
      "validation_rows": 1954,
      "input_sha256": {
        "train": "8abfac7c6107820270e9cdbfe6bdbb4ade5a37b0882c454db979c462df5e03ca",
        "validation": "331199cfff34de75959d4e9d0ead0aff8535d69ef079d5fb175cba6f3b43e97d"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": null,
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "official_zip",
        "url": "https://storage.googleapis.com/ai2-mosaic/public/socialiqa/socialiqa-train-dev.zip",
        "sha256": "ee073914a0fc33265cfbcfc50ec20df9b2e07809c3f420f599138d1f394ef5c3",
        "bytes": 2198056,
        "revision": null,
        "revision_basis": "Content-pinned archive: freshly fetched bytes match the historical ZIP; no Git revision applies."
      },
      "exporter_key": "social_iqa",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "e86f9aa59a13307e2860867c34371b21bbe9b5fc897e9c605fbc2ff95786a10e",
        "validation": "6e18f0c78febccd11b084d08c856bc6d1a83ecd3f4603f36a33a5f9d2de6a023"
      },
      "unique_sample_ids": {
        "train": 33410,
        "validation": 1954
      }
    },
    "arce": {
      "train_rows": 2251,
      "validation_rows": 570,
      "input_sha256": {
        "train": "e5d440429d63edaa29b7bc35968d5bd9f162156102ad5e3aa90b09fcfef8aed5",
        "validation": "20d1457a621b0ad6a166bb35fa0b7f363c9184eb54fc680e12100b843978ad35"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": "210d026faf9955653af8916fad021475a3f00453",
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "huggingface_parquet",
        "repository": "allenai/ai2_arc",
        "original_repository": "ai2_arc",
        "configuration": "ARC-Easy",
        "revision": "210d026faf9955653af8916fad021475a3f00453",
        "archived_metadata_sha256": "f36e9fd2172615e51446854a5a3705fbc3d830092c6741d38068cc37539f301b",
        "files": {
          "train": {
            "path": "ARC-Easy/train-00000-of-00001.parquet",
            "sha256": "b315db8a4be597dc7daa50a4e70d48dd7c990c32085629e6ccd8c926beaa80b5",
            "bytes": 330598
          },
          "validation": {
            "path": "ARC-Easy/validation-00000-of-00001.parquet",
            "sha256": "ed890ff1e4cef7a7140d3a30dcea3ed2c9d467c6458f447ad9ef0176d8dcbb74",
            "bytes": 86080
          }
        }
      },
      "exporter_key": "arc_easy",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "2df5e970b314ac8f30a27ab6abe0dcebfc39a42b260b6fd17bd7f1f847d5f06f",
        "validation": "41a3a0fe1751dfc86479e9ae38dafdc35dc31fe8cfc1ae7eedda529576c5be6c"
      },
      "unique_sample_ids": {
        "train": 2251,
        "validation": 570
      }
    },
    "winogrande": {
      "train_rows": 40398,
      "validation_rows": 1267,
      "input_sha256": {
        "train": "796969bbe3db8e8fa2fcde7bbfd05a1fe8f245c0165939cb110cf14ec06378cc",
        "validation": "bc6dc5a27ae4e063e3e4889f596df991d2a269f1c27ef9f6c0769ea7ff746a28"
      },
      "source_format": "archived CSV exports; not freshly downloaded Arrow",
      "upstream_revision": "01e74176c63542e6b0bcb004dcdea22d94fb67b5",
      "release_status": "content_matched_reference_reconstruction_verified",
      "source": {
        "kind": "huggingface_parquet",
        "repository": "allenai/winogrande",
        "original_repository": "winogrande",
        "configuration": "winogrande_xl",
        "revision": "01e74176c63542e6b0bcb004dcdea22d94fb67b5",
        "archived_metadata_sha256": "8ede0bea97fab79baffd054972171f9e36710a6435799d91126c50df9fab1d29",
        "files": {
          "train": {
            "path": "winogrande_xl/train-00000-of-00001.parquet",
            "sha256": "5d8c38ad12b9a6c88f79b6e00aaf0f40781f93d4f94816f6cff2b67625d67399",
            "bytes": 2058506
          },
          "validation": {
            "path": "winogrande_xl/validation-00000-of-00001.parquet",
            "sha256": "f9d914d1818c0aba98cabdf8af2bc4bd943c462b9c5e7062647e05c2ce3315d1",
            "bytes": 85928
          }
        }
      },
      "exporter_key": "winogrande",
      "reconstruction_script": "prepare_reference_datasets.py",
      "ordered_ids_sha256": {
        "train": "5c3bf73e0f4bdb38e1d592967cdf7b3176fbbdd155f0f5a3539248d23ca96a2f",
        "validation": "1dc2a74753777c8474d87f670672ec8a6b3bc962f88232b7f1f33172e2dc55ce"
      },
      "unique_sample_ids": {
        "train": 40398,
        "validation": 1267
      }
    }
  },
  "pending": [
    "Historical runs used different client counts and seeds; no single universal schedule is certified. The reference runtime now records explicit client/seed/round/step settings and fixed-context preprocessing per task. Its current contract does not establish that every old leaderboard score used those settings."
  ],
  "reference_runtime_contract": {
    "version": "legacy_fixed_context_correct_answer_position.v1",
    "guide": "integrations/pytorch/REFERENCE_MATRIX.md",
    "entrypoint": "integrations/pytorch/run_reference_matrix.py",
    "partition": "raw_row_index_mod_clients_before_filter",
    "required_run_settings": ["clients", "seed", "rounds", "local_steps"],
    "centralized_protocols": ["A", "B"],
    "default_evidence": "development_only"
  },
  "dataset_export_contract": {
    "version": "recovered-csv-export-20260421-v1",
    "original_exporter_sha256": "1d2e47f1e5632eeaa214b3ddb5997566d484e46950908075229b6a6ccbaabc72",
    "encoding": "UTF-8",
    "csv_line_terminator": "CRLF",
    "fields": [
      "sample_id",
      "question",
      "choices_json",
      "answer",
      "subject",
      "num_choices"
    ],
    "normalization": "CRLF/CR to LF; strip outer whitespace only; retain interior whitespace and source order.",
    "splits": "Source train and validation (dev for the two official ZIPs); no test split, shuffle, subsampling or token truncation at CSV export.",
    "identity": "Canonical row key: dataset_id/split/zero-based-row-index (e.g. hellaswag/validation/0). Preserve sample_id as source metadata; it can repeat even within a split. Ordered ID hash is SHA-256 of UTF-8 concatenated native sample_id + LF in CSV order.",
    "verification": "All 14 CSVs reconstructed from fresh pinned public sources and matched preserved reference hashes on 2026-10-19. Not proof of every historical result's input or a full training recipe."
  }
}