{
  "repository": "UnimeType/Transliteration-2.0-9B",
  "version": "2.6",
  "model_version": "2.6 (dataset_v5.1 12-language upgrade)",
  "release_tag": "v2.6",
  "protocol_version": "unimetype-convert-short-v3",
  "format": "MLX 6-bit affine, group size 64",
  "base_model": "Qwen/Qwen3.5-9B",
  "base_snapshot_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "base_source_sha256": {
    "model.safetensors-00001-of-00004.safetensors": "db6f444b43d318c92f360a13a25561a6a65b10c0631b8ed305a426dbaa6c380e",
    "model.safetensors-00002-of-00004.safetensors": "31c7d7e2dd5d207840b31cc59083c8f4c4718959149e0358c0364052bb9a0330",
    "model.safetensors-00003-of-00004.safetensors": "7ec36ba3a4176a44c3c0876ad80c56a2f70c84bf008d82e9501df642f17dadec",
    "model.safetensors-00004-of-00004.safetensors": "b62b0c4cd7e44edee103ee8f4fe225f246d5e768e07bfd5f25b63a8aa1fdd0c6"
  },
  "lora_adapter_sha256": "ed51a188f09bef51c82fba732e8c9c8990998175d63a24b2994f97b4dcbe43e4",
  "lora_hyperparameters": {
    "r": 32,
    "lora_alpha": 96.0,
    "scale": 3.0,
    "num_layers": 32,
    "fused_linear_modules": 248
  },
  "training_method": "LoRA SFT + Zero-Overlap Contrastive Preference Alignment (r=32, alpha=96, all 32 layers) on native unquantized BF16 Qwen/Qwen3.5-9B across 12 languages / 13 modes",
  "native_unquantized_base_precision": true,
  "training_dataset_public": false,
  "source_safetensors_repo": "UnimeType/Transliteration-2.0-9B-Safetensors",
  "standard_weight_sha256": {
    "model-00001-of-00004.safetensors": "800d3318a76943b7313231e4dbc64e34f4b5c94a180e10f65e2ab248847beb2a",
    "model-00002-of-00004.safetensors": "ab374a3bceae7c48d61701b7b3ec8034eccb3035083f4205f96e07468e55ad85",
    "model-00003-of-00004.safetensors": "3973ff5f69d3f9f7279258673ef22835d4d77974747b94f654b4ce03f1d3d321",
    "model-00004-of-00004.safetensors": "cc10ea28a11931916b5befcb6fc9e9314c78ae42e0cce19401f9b3a22faf8b78"
  },
  "published_weight_sha256": {
    "model-00001-of-00002.safetensors": "c369564008e7e976b241d6eb7b81351cb2eec96684d60388903e4fb0e7751307",
    "model-00002-of-00002.safetensors": "9390344fcd1d769b8a0fdf6342d33fd1876cbcb154ab0a47e88777e520c16562"
  },
  "evaluation": {
    "protocol": "unimetype-convert-short-v3",
    "dataset_generation": "dataset_v5.1_clean_12lang (0.00% eval overlap)",
    "cloud_reference": {
      "model": "deepseek/deepseek-v4.1-flash (system prompt ~2,800 tok)",
      "v4_shared_708": 0.9082,
      "core5_773": 0.903,
      "multilingual_12lang_1573": 0.932,
      "mst_bench_520_exact": 0.7346,
      "mst_bench_520_tolerant": 0.7481,
      "dogfood_30_exact": 0.5333,
      "dogfood_30_tolerant": 0.7667,
      "ood_unseen_20_probe_repair": 0.9
    },
    "models": {
      "0.8B": {
        "best_epoch": 3.0,
        "valid_loss": 0.263216,
        "v4_shared_708": 0.7232,
        "core5_773": 0.7115,
        "multilingual_12lang_1573": 0.6999,
        "mst_bench_520_exact": 0.4115,
        "mst_bench_520_tolerant": 0.4173,
        "dogfood_30_exact": 0.3667,
        "dogfood_30_tolerant": 0.3667,
        "ood_unseen_20_probe_repair": 0.45
      },
      "4B": {
        "best_epoch": 2.0,
        "valid_loss": 0.1105,
        "v4_shared_708": 0.8912,
        "core5_773": 0.8745,
        "multilingual_12lang_1573": 0.8811,
        "mst_bench_520_exact": 0.6327,
        "mst_bench_520_tolerant": 0.6385,
        "dogfood_30_exact": 0.5,
        "dogfood_30_tolerant": 0.5667,
        "ood_unseen_20_probe_repair": 0.75
      },
      "9B": {
        "best_epoch": 2.0,
        "valid_loss": 0.0799,
        "v4_shared_708": 0.9209,
        "core5_773": 0.9094,
        "multilingual_12lang_1573": 0.9154,
        "mst_bench_520_exact": 0.6558,
        "mst_bench_520_tolerant": 0.6731,
        "dogfood_30_exact": 0.7333,
        "dogfood_30_tolerant": 0.7667,
        "ood_unseen_20_probe_repair": 0.9
      }
    }
  }
}