{
  "schema_version": 1,
  "repo_id": "Nicholas0228/loop-distil-checkpoints",
  "weights_revision": "caa36badb2df9481350bb5945d1f7192971afa10",
  "source_code_commit": "bf8005f43a2be2b9ca0cc2c7fed3512bcc8664a0",
  "base_model": "tomg-group-umd/huginn-0125",
  "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
  "created_utc": "2026-10-09T20:30:17.370553+00:00",
  "checkpoint_count": 12,
  "logical_weights_bytes": 39296431104,
  "format": "loop-distil-recurrent-bf16-v1",
  "optimizer_included": false,
  "evaluation_scope": "128 repeatedly reused held-out GSM8K development questions; one training seed; automatic numeric answer matching, not a fresh test-set estimate.",
  "checkpoints": [
    {
      "id": "4loop-stage1-4000",
      "student_depth": 4,
      "teacher_depth": 32,
      "stage1_updates": 4000,
      "stage2_updates": 0,
      "checkpoint_step": 4000,
      "learning_rate": 2e-06,
      "comparison_report": "reports/four_loop_long_init_20261009/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 34,
        "strict_matches": 25,
        "cap_hits": 3,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "f99e853cc4ca2314032e7e354a43da477b087dc3723eef585606ed92cc245848",
      "source_master_sha256": "3fbc991e54a9ad09a8b5cd03ac7e50d9b092281a6615c74f96ea54d7195a9817",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_4000.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "7ffd5737aa8dfbf7af312c7330e907c78e75c762"
    },
    {
      "id": "4loop-stage1-3500",
      "student_depth": 4,
      "teacher_depth": 32,
      "stage1_updates": 3500,
      "stage2_updates": 0,
      "checkpoint_step": 3500,
      "learning_rate": 2e-06,
      "comparison_report": "reports/four_loop_long_init_20261009/comparison.json",
      "selection": "best observed development accuracy; selected after evaluation",
      "evaluation": {
        "questions": 128,
        "final_matches": 38,
        "strict_matches": 26,
        "cap_hits": 7,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "b5ed80020f3288b146534014079b6e9b82d549f11995ddcf50dd0884e1b6c242",
      "source_master_sha256": "6a9c8dd2ca09e4bb80aa27166ee95e07abf43938e5a71423c376a10a5ad698a3",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_3500.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "8b9e683d76a7aaac3be929568aefe3ae34a55421"
    },
    {
      "id": "4loop-stage1-2000",
      "student_depth": 4,
      "teacher_depth": 32,
      "stage1_updates": 2000,
      "stage2_updates": 0,
      "checkpoint_step": 2000,
      "learning_rate": 2e-06,
      "comparison_report": "reports/four_loop_long_init_20261009/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 28,
        "strict_matches": 21,
        "cap_hits": 5,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "11443bbd5815d9ced04b97bac963c0fb71a2804af62cac88640deac0eb9db567",
      "source_master_sha256": "9126e5b82feb48af5d90eb06288afecc5e09fb27f388360cfe55529c4f9a72f5",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_2000.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "20491d260c900a74dd2110acd5001e2830f5286c"
    },
    {
      "id": "4loop-stage2-direct-500",
      "student_depth": 4,
      "teacher_depth": 32,
      "stage1_updates": 2000,
      "stage2_updates": 500,
      "checkpoint_step": 500,
      "learning_rate": 2e-07,
      "comparison_report": "reports/four_loop_extend_20261009/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 32,
        "strict_matches": 26,
        "cap_hits": 3,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "7b01376b3b178afe77791fff292d5d3eedc2e6d9fbfb75b4d15db56a6d0a73da",
      "source_master_sha256": "1b7549a6766c736ec793dbdf6196146365876765c0388b240782bc6c9dd7725c",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_0500.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "8332597d21f85f332f7d0161fe512a763f84b2da"
    },
    {
      "id": "4loop-stage2-proposal-500",
      "student_depth": 4,
      "teacher_depth": 32,
      "stage1_updates": 2000,
      "stage2_updates": 500,
      "checkpoint_step": 500,
      "learning_rate": 2e-07,
      "comparison_report": "reports/four_loop_extend_20261009/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 31,
        "strict_matches": 22,
        "cap_hits": 1,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "ac2856418979243e31cad3368c39bc469f52961e24e06d707142056208fc9ee3",
      "source_master_sha256": "da7be0948a46eca7028b597cec5b8486d283a21a3795e89cc5ce18616d5e930c",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_0500.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "66559c76a935d316dc6855cffbfce47a7418c643"
    },
    {
      "id": "8loop-stage1-1000",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 0,
      "checkpoint_step": 1000,
      "learning_rate": 2e-06,
      "comparison_report": "reports/intermediate_low_lr_20261008/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 46,
        "strict_matches": 36,
        "cap_hits": 2,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "a3112f972f78ebc13338a9296deb0155bde7f197eb69fbae6997d38d443fc44a",
      "source_master_sha256": "b1955400c40d652259c924212dfe6b407ef5eb95f4332d4894e7b321f7bb4883",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "warm_start.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "e4fd40a9bd5808d714aa151dd94f3b783738b652"
    },
    {
      "id": "8loop-stage2-direct-low-lr-100",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 100,
      "checkpoint_step": 100,
      "learning_rate": 2e-07,
      "comparison_report": "reports/intermediate_low_lr_20261008/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 48,
        "strict_matches": 36,
        "cap_hits": 3,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "7f14d5e48c387ed6913f77fd33e48c0f9767985a2acd1ae9d725e245ba28ab74",
      "source_master_sha256": "d7e240de21ab0e38565da0e3a1d1802c29e840cae6c30af1f5ec9d65af555934",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "best_fixed_kl.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "c70766cf6ae2c6096930c071a67f73fb0f4b1651"
    },
    {
      "id": "8loop-stage2-proposal-low-lr-100",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 100,
      "checkpoint_step": 100,
      "learning_rate": 2e-07,
      "comparison_report": "reports/intermediate_low_lr_20261008/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 55,
        "strict_matches": 42,
        "cap_hits": 4,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "acb347b832fd02c38c8a72a88497aa3df70b50bfa7f3b3c73a0bf8dfa0f4572a",
      "source_master_sha256": "8fbe2878e0308f3e2399885f0f8ae96943cff77404e5ed3bf31f8b849894d125",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "latest.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "6e6f1e8ffe5a60bd2b6fdfdaaf5acaea75435fdc"
    },
    {
      "id": "8loop-stage2-direct-high-lr-100",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 100,
      "checkpoint_step": 100,
      "learning_rate": 2e-06,
      "comparison_report": "reports/intermediate_followup_20261008/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 32,
        "strict_matches": 23,
        "cap_hits": 5,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "189a82e3c55d60c3155c132cac5aa3ee360ceda683fc7007528290ee09701ef6",
      "source_master_sha256": "53fb50ed90e03ba2d9808e32e53c14d2e4d2f14e764c02174ab4821924bc0c13",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "latest.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "2a542a8194eccd7e0345ff8d0b4ee14c9bccb226"
    },
    {
      "id": "8loop-stage2-proposal-high-lr-100",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 100,
      "checkpoint_step": 100,
      "learning_rate": 2e-06,
      "comparison_report": "reports/intermediate_followup_20261008/comparison.json",
      "selection": "predeclared endpoint",
      "evaluation": {
        "questions": 128,
        "final_matches": 8,
        "strict_matches": 5,
        "cap_hits": 11,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "d1897b8c33c11ee0bc3854d116696aa9efd2d450df98e6e6e660d74a1a986c22",
      "source_master_sha256": "e6778eb092220baf8a6591cedd83f8c404d61b39b818edd7551456550851b7d2",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "latest.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "d1b9e0a69847e5b16c92198b1c0e5d607edf1c15"
    },
    {
      "id": "8loop-stage2-direct-low-lr-300",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 300,
      "checkpoint_step": 300,
      "learning_rate": 2e-07,
      "comparison_report": "reports/continuation_500_20261008/comparison.json",
      "selection": "last fully evaluated matched checkpoint before user stop",
      "evaluation": {
        "questions": 128,
        "final_matches": 47,
        "strict_matches": 36,
        "cap_hits": 2,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "e5ee42530f6fbcee7502bb89bf267e626f4f110af0ee65bc973aac1fce596113",
      "source_master_sha256": "3ebff68fca558ba6512fe9250963ccca32b24d994cc05cd1a29efc0914c01035",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_0300.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "ba864e42d6b76679a080b3c6341a4ec2f3642c2d"
    },
    {
      "id": "8loop-stage2-proposal-low-lr-300",
      "student_depth": 8,
      "teacher_depth": 32,
      "stage1_updates": 1000,
      "stage2_updates": 300,
      "checkpoint_step": 300,
      "learning_rate": 2e-07,
      "comparison_report": "reports/continuation_500_20261008/comparison.json",
      "selection": "last fully evaluated matched checkpoint before user stop",
      "evaluation": {
        "questions": 128,
        "final_matches": 48,
        "strict_matches": 34,
        "cap_hits": 2,
        "split": "held-out GSM8K development",
        "generation_cap": 256
      },
      "prefix_policy": "cached teacher32-generated answer prefixes",
      "format": "loop-distil-recurrent-bf16-v1",
      "base_model": "tomg-group-umd/huginn-0125",
      "base_revision": "bb6621b65e90b6a4b9b29ef88dc83866d450470c",
      "weights_sha256": "67197f0bf7b28d2b071b42e6cb56f2a7b71768845173745144187246ea4dc787",
      "source_master_sha256": "a7598488d52fb8818212b35939bfd9ccff4ed847d9879f3322ab5750f151fe17",
      "tensor_count": 37,
      "parameters": 1637349120,
      "weights_bytes": 3274702592,
      "dtype": "bfloat16",
      "exact_bf16_compute_roundtrip": true,
      "optimizer_included": false,
      "source_checkpoint_filename": "step_0300.pt",
      "note": "Full replacements for adapter/core tensors, not additive deltas or LoRA. Combine with the pinned base. Original FP32 masters and Adam states remain in local recovery checkpoints.",
      "checkpoint_revision": "caa36badb2df9481350bb5945d1f7192971afa10"
    }
  ]
}