{
  "data": {
    "dev": "/content/jevlet_data/mixture_v6/dev.jsonl",
    "train": "/content/jevlet_data/mixture_v6/train.jsonl",
    "vault": "/content/jevlet_data/mixture_v6/vault/vault.jsonl"
  },
  "device": "auto",
  "evaluation": {
    "batch_size": 64,
    "benchmark_repeats": 5,
    "max_examples": 6000,
    "permutation_examples": 64
  },
  "model": {
    "attention_topology": "block_bidir",
    "backbone": "BAAI/bge-small-en-v1.5",
    "decision_head": "pointer",
    "family": "pretrained",
    "gradient_checkpointing": false,
    "max_option_tokens": 40,
    "max_packed_len": 1024,
    "max_question_tokens": 64,
    "max_state_tokens": 224,
    "option_pool": "mean"
  },
  "notes": "Same 1.6M-example budget as pretrained_daily_v6.json (100k x 16) at effective batch 64: 25k steps, learning rates scaled by sqrt(4), warmup and snapshots at the same example counts. jevlet.colab fits batch_size x gradient_accumulation to the GPU and keeps their product at effective_batch.",
  "provenance": {
    "base_config": "configs/pretrained_daily_v6_colab.json",
    "created": 1790338776.0874045,
    "git_commit": "9d62c46989e965e0d823491e670dc01ed12a345e"
  },
  "seed": 1337,
  "sessions": [
    {
      "gpu": {
        "batch_size": 64,
        "bf16": true,
        "memory_gb": 79.3,
        "name": "NVIDIA A100-SXM4-80GB",
        "workers": 8
      },
      "resume": false,
      "time": 1790338782.2884605
    },
    {
      "gpu": {
        "batch_size": 64,
        "bf16": true,
        "memory_gb": 79.3,
        "name": "NVIDIA A100-SXM4-80GB",
        "workers": 8
      },
      "resume": true,
      "time": 1790352643.7766407
    }
  ],
  "training": {
    "amp_dtype": "bf16",
    "backbone_learning_rate": 0.0001,
    "batch_size": 64,
    "effective_batch": 64,
    "gradient_accumulation": 1,
    "keep_steps": [
      2500,
      6250,
      12500
    ],
    "learning_rate": 0.002,
    "log_every_steps": 50,
    "loss": "ce_brier",
    "lr_schedule": "cosine",
    "max_grad_norm": 1.0,
    "max_steps": 25000,
    "min_lr_ratio": 0.05,
    "mirror_dir": "/content/drive/MyDrive/Jevlet/runs/v6",
    "num_workers": 8,
    "resume_from": "/content/runs/v6/resume.pt",
    "save_every_steps": 500,
    "warmup_steps": 500,
    "weight_decay": 0.01
  }
}