[
  {
    "assistant_tokens_recorded": 4440608,
    "bundle": "step_0002000_f3d1ffd185f9",
    "input_tokens_recorded": 29271514,
    "manifest_sha256": "f9273450799963b2fd8e61a5e9cb6892be7ba4a2308849623e288302a687a1f7",
    "optimizer": "Fresh AdamW moments at each SFT stage",
    "parent_manifest": null,
    "plan": "configs/sft1/plan.json",
    "prose_replay": {
      "presentations": 32000,
      "tokens": 16384000
    },
    "source_pretraining_tokens": 42001367040,
    "stage": "sft1",
    "steps": 2000,
    "supervised": {
      "components": {
        "constraints": 12400,
        "dialogue": 16000,
        "general": 9600,
        "grounded": 22000,
        "knowledge": 8000,
        "rewrite": 2000,
        "skills": 24000,
        "social": 2000
      },
      "maximum_serialized_tokens": 1385,
      "presentations": 96000,
      "subsets": {
        "sciq": 8000,
        "smol": 12000,
        "smol_constraints": 4000,
        "smol_rewrite": 2000,
        "social": 2000,
        "squad": 22000,
        "verified_constraints": 8400,
        "verified_dialogue": 13600,
        "verified_skills": 24000
      },
      "unique_rows": 48000
    },
    "training_sources": [
      {
        "label": "smol",
        "license": "apache-2.0",
        "parquet_urls": [
          "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00000-of-00004.parquet",
          "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00001-of-00004.parquet",
          "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00002-of-00004.parquet",
          "https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk/resolve/f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc/data/train-00003-of-00004.parquet"
        ],
        "repo": "HuggingFaceTB/smol-smoltalk",
        "revision": "f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc",
        "split": "train"
      },
      {
        "label": "social",
        "license": "see upstream component license",
        "parquet_urls": [
          "https://huggingface.co/datasets/HuggingFaceTB/smoltalk/resolve/5feaf2fd3ffca7c237fc38d1861bc30365d48ffa/data/everyday-conversations/train-00000-of-00001.parquet"
        ],
        "repo": "HuggingFaceTB/smoltalk",
        "revision": "5feaf2fd3ffca7c237fc38d1861bc30365d48ffa",
        "split": "train"
      },
      {
        "label": "squad",
        "license": [
          "cc-by-sa-4.0"
        ],
        "parquet_urls": [
          "https://huggingface.co/datasets/rajpurkar/squad_v2/resolve/3ffb306f725f7d2ce8394bc1873b24868140c412/squad_v2/train-00000-of-00001.parquet"
        ],
        "repo": "rajpurkar/squad_v2",
        "revision": "3ffb306f725f7d2ce8394bc1873b24868140c412",
        "split": "train"
      },
      {
        "label": "sciq",
        "license": "cc-by-nc-3.0",
        "parquet_urls": [
          "https://huggingface.co/datasets/allenai/sciq/resolve/2c94ad3e1aafab77146f384e23536f97a4849815/data/train-00000-of-00001.parquet"
        ],
        "repo": "allenai/sciq",
        "revision": "2c94ad3e1aafab77146f384e23536f97a4849815",
        "split": "train"
      }
    ]
  },
  {
    "actual_sources_note": "Fresh tasks are authored/computed; replay is selected from stage-1 training. Inherited plan.sources and corpus manifest describe broader/stale configuration, not additional fetched sources.",
    "assistant_tokens_recorded": 279606,
    "bundle": "step_0000256_6074b98fe66b",
    "fresh": {
      "components": {
        "constraints": 1638,
        "dialogue": 410,
        "grounded": 410,
        "skills": 1638
      },
      "maximum_serialized_tokens": 115,
      "presentations": 4096,
      "subsets": {
        "unspecified": 4096
      },
      "unique_rows": 4096
    },
    "fresh_file_sha256": "0a90cbb82b2587f20026dd21c9d1697ab37016c7714e19caade3ce4c94856dcc",
    "input_tokens_recorded": 816428,
    "manifest_sha256": "dcab4ad8e4484ad1cc4b32ffbc6352b6a6ff9ea0a4ddccdb8d367d23db098a3f",
    "optimizer": "Fresh AdamW moments at each SFT stage",
    "parent_manifest": "f9273450799963b2fd8e61a5e9cb6892be7ba4a2308849623e288302a687a1f7",
    "plan": "configs/sft2/plan.json",
    "replay": {
      "components": {
        "constraints": 512,
        "dialogue": 512,
        "general": 768,
        "grounded": 512,
        "knowledge": 768,
        "rewrite": 256,
        "skills": 512,
        "social": 256
      },
      "maximum_serialized_tokens": 512,
      "presentations": 4096,
      "subsets": {
        "sciq": 768,
        "smol": 781,
        "smol_constraints": 162,
        "smol_rewrite": 256,
        "social": 256,
        "squad": 512,
        "verified_constraints": 350,
        "verified_dialogue": 499,
        "verified_skills": 512
      },
      "unique_rows": 4096
    },
    "replay_file_sha256": "3b7897ad24c215c309b1a693dc6b3606783c678e82f7d595bdc3f9f97d8797ab",
    "source_pretraining_tokens": 42001367040,
    "stage": "sft2",
    "steps": 256
  },
  {
    "actual_sources_note": "Fresh tasks are authored/computed; replay is selected from stage-1 training. Inherited plan.sources and corpus manifest describe broader/stale configuration, not additional fetched sources.",
    "assistant_tokens_recorded": 120629,
    "bundle": "step_0000128_5937a6def881",
    "fresh": {
      "components": {
        "abstention": 188,
        "dialogue": 306,
        "grounded": 324,
        "reasoning": 1230
      },
      "maximum_serialized_tokens": 137,
      "presentations": 2048,
      "subsets": {
        "unspecified": 2048
      },
      "unique_rows": 2048
    },
    "fresh_file_sha256": "c32660b92756608b1c949730c61595dc78d29b0afb3695caf6cfc8397d4acbf5",
    "input_tokens_recorded": 411448,
    "manifest_sha256": "2118c7446c412a4ee7a4b7a356a211ad91574983e213da45980707b5fa1ca124",
    "optimizer": "Fresh AdamW moments at each SFT stage",
    "parent_manifest": "dcab4ad8e4484ad1cc4b32ffbc6352b6a6ff9ea0a4ddccdb8d367d23db098a3f",
    "plan": "configs/sft3/plan.json",
    "replay": {
      "components": {
        "constraints": 246,
        "dialogue": 317,
        "general": 352,
        "grounded": 226,
        "knowledge": 413,
        "rewrite": 130,
        "skills": 254,
        "social": 110
      },
      "maximum_serialized_tokens": 512,
      "presentations": 2048,
      "subsets": {
        "sciq": 413,
        "smol": 365,
        "smol_constraints": 73,
        "smol_rewrite": 130,
        "social": 110,
        "squad": 226,
        "verified_constraints": 173,
        "verified_dialogue": 304,
        "verified_skills": 254
      },
      "unique_rows": 2048
    },
    "replay_file_sha256": "3b7897ad24c215c309b1a693dc6b3606783c678e82f7d595bdc3f9f97d8797ab",
    "source_pretraining_tokens": 42001367040,
    "stage": "sft3",
    "steps": 128
  }
]