{
  "title": "HYDRA Training Logs — 1.5B & 7B",
  "export_date": "2026-09-28",
  "report_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/reports/HYDRA-Training-Logs-1-5B-7B--VmlldzoxODAxNzM2OQ",
  "selection_rule": "Latest finished run within each model/configured stage, with complete exported loss coverage. 7B Stage III explicitly confirmed by the author.",
  "checkpoint_provenance_note": "The five automatically selected runs have not been independently matched to the paper final checkpoints.",
  "metrics": [
    "step_loss_ntp",
    "step_loss_flow"
  ],
  "logging_interval_steps": 50,
  "smoothing": 0,
  "runs": [
    {
      "model_size": "1.5B",
      "stage": 1,
      "run_id": "a06r4cgg",
      "original_name": "unify-tokenizer-1.5b-stage-1-kl-2.0-nf",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/a06r4cgg",
      "created_utc": "2026-01-14T07:39:31.000Z",
      "original_state": "finished",
      "selection": "latest_finished_with_complete_loss_coverage",
      "records": 3000,
      "first_step": 50,
      "last_step": 150000,
      "csv_path": "logs/HYDRA_1.5B_stage1.csv",
      "sha256": "2dfa6355a5908fc730183601583716e2ef66246ffb4eac0237387deb13800939",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "2000",
        "lr_scheduler.scheduler": "constant_with_warmup",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00005",
        "optimizer.params.learning_rate_showo": "0.0001",
        "optimizer.params.learning_rate_ve": "0.00005",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0",
        "training.batch_size_mmu": "1",
        "training.batch_size_t2i": "3",
        "training.cond_dropout_prob": "0.1",
        "training.ema_update_every": "5",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "150000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "2",
        "training.ntp_coeff": "0.8",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true"
      }
    },
    {
      "model_size": "1.5B",
      "stage": 2,
      "run_id": "bd8o2c5q",
      "original_name": "unify-tokenizer-1.5b-stage-2-kl-2.0-nf",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/bd8o2c5q",
      "created_utc": "2026-01-19T03:35:30.000Z",
      "original_state": "finished",
      "selection": "latest_finished_with_complete_loss_coverage",
      "records": 2000,
      "first_step": 50,
      "last_step": 100000,
      "csv_path": "logs/HYDRA_1.5B_stage2.csv",
      "sha256": "bfb25edcf98d4a77a4cfbdbcd76d1f75bd4668fe40374bdf70c80dade91370b9",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "1000",
        "lr_scheduler.scheduler": "constant_with_warmup",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00002",
        "optimizer.params.learning_rate_showo": "0.00002",
        "optimizer.params.learning_rate_ve": "0.000002",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0.01",
        "training.batch_size_mmu": "2",
        "training.batch_size_t2i": "2",
        "training.cond_dropout_prob": "0.1",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "100000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "2",
        "training.ntp_coeff": "1",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true",
        "optimizer.params.learning_rate_dit": "0.00005",
        "training.stage": "pre-training-1-5"
      }
    },
    {
      "model_size": "1.5B",
      "stage": 3,
      "run_id": "hhfvjzd5",
      "original_name": "unify-tokenizer-1.5b-stage-3-kl-1.0-nf",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/hhfvjzd5",
      "created_utc": "2026-01-22T05:22:18.000Z",
      "original_state": "finished",
      "selection": "latest_finished_with_complete_loss_coverage",
      "records": 400,
      "first_step": 50,
      "last_step": 20000,
      "csv_path": "logs/HYDRA_1.5B_stage3.csv",
      "sha256": "2e5bf55572ba00e7e72b0a84f0fecc121f4e9c0a5291f3bc4ea58b0401829488",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "10",
        "lr_scheduler.scheduler": "cosine",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00002",
        "optimizer.params.learning_rate_showo": "0.00002",
        "optimizer.params.learning_rate_ve": "0.000002",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0.01",
        "training.batch_size_mmu": "2",
        "training.batch_size_t2i": "2",
        "training.cond_dropout_prob": "0.1",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "20000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "1",
        "training.ntp_coeff": "1",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true",
        "lr_scheduler.params.min_lr": "0.000001",
        "training.stage": "tuning"
      }
    },
    {
      "model_size": "7B",
      "stage": 1,
      "run_id": "u2i0d0dn",
      "original_name": "unify-tokenizer-7b-stage-1-kl",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/u2i0d0dn",
      "created_utc": "2026-01-15T11:00:18.000Z",
      "original_state": "finished",
      "selection": "latest_finished_with_complete_loss_coverage",
      "records": 3000,
      "first_step": 50,
      "last_step": 150000,
      "csv_path": "logs/HYDRA_7B_stage1.csv",
      "sha256": "5d4ff2cfcc121e66349e21990b3ab35865a44dc8f6bd101d3c8105a500c6fca3",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "2000",
        "lr_scheduler.scheduler": "constant_with_warmup",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00005",
        "optimizer.params.learning_rate_showo": "0.0001",
        "optimizer.params.learning_rate_ve": "0.00005",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0",
        "training.batch_size_mmu": "1",
        "training.batch_size_t2i": "3",
        "training.cond_dropout_prob": "0.1",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "150000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "2",
        "training.ntp_coeff": "0.8",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true"
      }
    },
    {
      "model_size": "7B",
      "stage": 2,
      "run_id": "8o9n0trm",
      "original_name": "unify-tokenizer-7b-stage-2-kl",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/8o9n0trm",
      "created_utc": "2026-01-20T09:12:43.000Z",
      "original_state": "finished",
      "selection": "latest_finished_with_complete_loss_coverage",
      "records": 2000,
      "first_step": 50,
      "last_step": 100000,
      "csv_path": "logs/HYDRA_7B_stage2.csv",
      "sha256": "d7d53334e7418d9fa46ba51826f9193adc0b871fb5cfddedaa304ead4b024ced",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "1000",
        "lr_scheduler.scheduler": "constant_with_warmup",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00002",
        "optimizer.params.learning_rate_showo": "0.00002",
        "optimizer.params.learning_rate_ve": "0.000002",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0.01",
        "training.batch_size_mmu": "2",
        "training.batch_size_t2i": "2",
        "training.cond_dropout_prob": "0.1",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "100000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "1",
        "training.ntp_coeff": "1",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true",
        "optimizer.params.learning_rate_dit": "0.00005"
      }
    },
    {
      "model_size": "7B",
      "stage": 3,
      "run_id": "375bggda",
      "original_name": "unify-tokenizer-7b-stage-3-kl-1.0-nf",
      "source_url": "https://wandb.ai/2223453260-uestc/Unify-tokenizer-new-kl/runs/375bggda",
      "created_utc": "2026-01-26T14:27:34.000Z",
      "original_state": "finished",
      "selection": "author_confirmed",
      "records": 400,
      "first_step": 50,
      "last_step": 20000,
      "csv_path": "logs/HYDRA_7B_stage3.csv",
      "sha256": "f59112c7880b24f6c66544d3c29560ec07aaf8d0b7355e5434389ed3ccdd37bf",
      "config_from_run": {
        "lr_scheduler.params.warmup_steps": "10",
        "lr_scheduler.scheduler": "cosine",
        "optimizer.params.beta1": "0.9",
        "optimizer.params.beta2": "0.999",
        "optimizer.params.epsilon": "1e-8",
        "optimizer.params.learning_rate_proj": "0.00002",
        "optimizer.params.learning_rate_showo": "0.00002",
        "optimizer.params.learning_rate_ve": "0.000002",
        "optimizer.params.scale_lr": "false",
        "optimizer.params.weight_decay": "0.01",
        "training.batch_size_mmu": "2",
        "training.batch_size_t2i": "2",
        "training.cond_dropout_prob": "0.1",
        "training.enable_tf32": "true",
        "training.flow_coeff": "1",
        "training.gradient_accumulation_steps": "1",
        "training.label_smoothing": "0",
        "training.max_grad_norm": "1",
        "training.max_train_steps": "20000",
        "training.mixed_precision": "bf16",
        "training.noise_shift": "1",
        "training.ntp_coeff": "1",
        "training.seed": "10086",
        "training.und_max_t0": "1",
        "training.use_torch_compile": "true",
        "lr_scheduler.params.min_lr": "0.000001",
        "training.stage": "tuning"
      }
    }
  ],
  "image_export_date": "2026-09-29",
  "image_metric": "Generated Validation Images",
  "coverage": [
    {
      "stage": 1,
      "model": "1.5B",
      "loss_rows": 3000,
      "loss_last_step": 150000,
      "image_steps": [
        2000,
        4000,
        6000,
        8000,
        10000,
        12000,
        14000,
        16000,
        18000,
        20000,
        22000,
        24000,
        26000,
        28000,
        30000,
        32000,
        34000,
        36000,
        38000,
        40000,
        42000,
        44000,
        46000,
        48000,
        50000,
        52000,
        54000,
        56000,
        58000,
        60000,
        62000,
        64000,
        66000,
        68000,
        70000,
        72000,
        74000,
        76000,
        78000,
        80000,
        82000,
        84000,
        86000,
        88000,
        90000,
        92000,
        94000,
        96000,
        98000,
        100000,
        102000,
        104000,
        106000,
        108000,
        110000,
        112000,
        114000,
        116000,
        118000,
        120000,
        122000,
        124000,
        126000,
        128000,
        130000,
        132000,
        134000,
        136000,
        138000,
        140000,
        142000,
        144000,
        146000,
        148000,
        150000
      ],
      "image_count": 225,
      "samples_per_checkpoint": 3
    },
    {
      "stage": 2,
      "model": "1.5B",
      "loss_rows": 2000,
      "loss_last_step": 100000,
      "image_steps": [
        10000,
        20000,
        30000,
        40000,
        50000,
        60000,
        70000,
        80000,
        90000,
        100000
      ],
      "image_count": 20,
      "samples_per_checkpoint": 2
    },
    {
      "stage": 3,
      "model": "1.5B",
      "loss_rows": 400,
      "loss_last_step": 20000,
      "image_steps": [
        5000,
        10000,
        15000,
        20000
      ],
      "image_count": 8,
      "samples_per_checkpoint": 2
    },
    {
      "stage": 1,
      "model": "7B",
      "loss_rows": 3000,
      "loss_last_step": 150000,
      "image_steps": [
        10000,
        20000,
        30000,
        40000,
        50000,
        60000,
        70000,
        80000,
        90000,
        100000,
        110000,
        120000,
        130000,
        140000,
        150000
      ],
      "image_count": 45,
      "samples_per_checkpoint": 3
    },
    {
      "stage": 2,
      "model": "7B",
      "loss_rows": 2000,
      "loss_last_step": 100000,
      "image_steps": [
        10000,
        20000,
        30000,
        40000,
        50000,
        60000,
        70000,
        80000,
        90000,
        100000
      ],
      "image_count": 20,
      "samples_per_checkpoint": 2
    },
    {
      "stage": 3,
      "model": "7B",
      "loss_rows": 400,
      "loss_last_step": 20000,
      "image_steps": [
        10000,
        20000
      ],
      "image_count": 4,
      "samples_per_checkpoint": 2
    }
  ],
  "image_manifest": "image_manifest.json"
}