{
  "paper": {
    "arxiv_id": "2412.09871",
    "title": "Byte Latent Transformer: Patches Scale Better Than Tokens",
    "url": "https://arxiv.org/abs/2412.09871",
    "organization": "Meta FAIR"
  },
  "dataset": {
    "name": "WikiText-2",
    "train_tokens": 120000,
    "validation_tokens": 8000,
    "test_tokens": 8000,
    "narrative_tokens": 120000,
    "instruction_train": 320,
    "instruction_validation": 64,
    "preference_validation": 64,
    "reasoning_validation": 0,
    "benchmark_suite": "core",
    "vocab_size": 512,
    "tokenizer": "data/llm-evolution/wikitext-narrative-bpe-512.json"
  },
  "setup": {
    "seed": 42,
    "steps_per_variant": 12,
    "dimensions": 64,
    "layers": 2,
    "sequence_length": 64,
    "same_tokens_optimizer_and_budget": true,
    "evolve_architecture": "blt",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 6.0430581122636795,
    "perplexity": 421.17908035954804,
    "instruction_loss": 6.0724897384643555,
    "primary": -6.953931573033333,
    "public_composite": -6.953931573033333,
    "composite_loss": 6.953931573033333,
    "fitness": -6.953931573033333,
    "initial_loss": 6.179897594451904,
    "final_loss": 6.145263195037842,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139584,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {},
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "method": {
    "name": "blt",
    "lm_loss": 6.021533235907555,
    "perplexity": 412.21012682777763,
    "instruction_loss": 6.0578306913375854,
    "primary": -6.930207839608192,
    "public_composite": -6.930207839608192,
    "composite_loss": 6.930207839608192,
    "fitness": -6.930207839608192,
    "initial_loss": 6.172091627120972,
    "final_loss": 6.139454650878906,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139649,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "tokenizer_free_reference": "dynamic byte patches",
      "observed_patch_boundary_rate": 0.8169800639152527,
      "expanded_back_to_byte_targets": 1.0
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": -0.3561917816484779,
    "perplexity_percent": -2.1294869451051275
  },
  "paper_results": {
    "tokenizer_free": true,
    "dynamic_entropy_patching": true
  },
  "scope": "实际执行可学习 surprisal boundary、相邻低熵位置 latent patch 共享并展开回原目标；当前 evaluator 的 512-symbol tokenizer 只近似 byte alphabet，未复刻 8B BLT 与完整 local encoder/decoder。",
  "schema_version": 2,
  "manifest_ref": "reproduction:blt",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2412.09871-blt/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}