{
  "arguments": {
    "backbone": "lfm2-vl",
    "attention": "causal",
    "label": "vl-a",
    "data": "data/snake-v1",
    "out": "/content/qevd-runs/vl-a",
    "epochs": 4,
    "batch": 8,
    "accum": 2,
    "lr": 0.0001,
    "lora_rank": 64,
    "seed": 42,
    "max_length": 1024,
    "max_state": 384,
    "eval_batch": 16,
    "device": "cuda",
    "limit_train": 0,
    "eval_limit": 0,
    "latency_items": 30,
    "latency_repeats": 10,
    "latency_questions": [
      1,
      4,
      16
    ],
    "skip_snake": false,
    "snake_seeds": 20,
    "snake_max_steps": 500,
    "checkpointing": false
  },
  "load_seconds": 23.635210275650024,
  "dataset_manifest_sha256": "f41151c68d6465b90bb8ea66ca0ea8611a6b37ed48596ba9484abc8ef4deda33",
  "questions": {
    "train": 17892,
    "calibration": 1120,
    "development": 1380
  },
  "state_truncations": {
    "train": 4,
    "calibration": 0,
    "development": 1
  },
  "tokens_per_question": {
    "train": {
      "n": 17892,
      "p50": 373.0,
      "p95": 377,
      "mean": 311.1713056114465
    },
    "calibration": {
      "n": 1120,
      "p50": 292.5,
      "p95": 377,
      "mean": 256.09732142857143
    },
    "development": {
      "n": 1380,
      "p50": 185.5,
      "p95": 377,
      "mean": 246.50797101449277
    }
  },
  "trainable_parameters": 24510464,
  "total_parameters": 473229312,
  "backbone": {
    "repo": "LiquidAI/LFM2.5-VL-450M",
    "revision": "fc6221ca597f3315e4f82fc2df606783267b34ba",
    "family": "lfm2",
    "kind": "lfm2_vl",
    "lora_targets": [
      "q_proj",
      "k_proj",
      "v_proj",
      "out_proj",
      "in_proj",
      "w1",
      "w2",
      "w3"
    ]
  },
  "hardware": "NVIDIA A100-SXM4-40GB",
  "packages": {
    "torch": "2.11.0+cu128",
    "transformers": "5.17.0",
    "peft": "0.21.0"
  },
  "source_sha256": {
    "__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
    "colab_collect.py": "799bd4c926d66a02c3a45c7c5c1f63a8b37c2358394d5e2a77c6fcd45d141314",
    "colab_djev.py": "43ce7057f9ec44281665ed34d1cf81e37bb39ebd02de13b75ff6ccb30626d4f1",
    "colab_extract.py": "602229803f6ebbbf672319552bf4d541732e5c02c4d95a830a7b7d11debbba43",
    "colab_launch.py": "24649c2b4ac4f7df12745a31cb2d1502ba26c189cda4bb118db42b545284249c",
    "colab_poll.py": "a1290383a131fd6f24fab883e935db2f57a8a1894610456a999aa9104eeae40a",
    "colab_refresh.py": "c814a3ebde100c39c692012b258f66175b3dd3921fd2025132345e4283571856",
    "colab_setup.py": "96fc394ab2ec0cd570ae4341d56a7e46dae5f7cf31c283d22c3ee576eff00526",
    "djev_zeroshot.py": "5f7c5f8051d3e96549da875fb498a575f419f91c46ca0a13d6f28bc61d45a541",
    "mlx_latency.py": "387453e4841977a04cceb4a1cdfccf0f43cc2998354b66c8a46f8b0ed3679342",
    "mlx_qev_adapters.py": "b3051d94d054cbd39de007093bc3ecd9e025e6a797dec54f4fc2a97d6cc79c0a",
    "mm_probe.py": "a39f30fb7ba2a7ad4b122156f79d9f385ce931dea8c96e0fc2eb848de9c2079e",
    "order_probe.py": "243d22c8b79c966641e8d7d09365d50b2ec2b0f9e6765e77fa0a8f3151994c45",
    "package.py": "0d5d66c9f9522c101ab11e69c20c04728faf4199874818a5981ffb99c03c4fde",
    "paired_tests.py": "7d6c954ad16de9dee1c6ec4f2bcdddbba50026d41f6ee3fea699bb3b3c89e4ef",
    "publish_archive.py": "c3fe4d5db33802cb357e99ab53e20cbdd1998c2f67f1827f0e66c03502e7dd4e",
    "qevd.py": "b59eaafdb0e6f38bc2a1040c485f40352a369ee7d75e4fb88c84ebb6bd712f2d",
    "report.py": "df2c4d8286d00b6ad0faa6024c90253734541fd9b5412297528f531fa0a4c974",
    "train_qevd.py": "426b410ae321f938306f7416a2e351666cce925629af0e7be766e602ca7011a0"
  },
  "objective": "supervised cross-entropy over candidate options (same recipe for every arm)"
}