{
  "paper": {
    "arxiv_id": "2502.13189",
    "title": "MoBA: Mixture of Block Attention for Long-Context LLMs",
    "url": "https://arxiv.org/abs/2502.13189",
    "organization": "Moonshot AI"
  },
  "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": "moba",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 6.043058142066002,
    "perplexity": 421.17909291166296,
    "instruction_loss": 6.0724897384643555,
    "primary": -6.953931602835655,
    "public_composite": -6.953931602835655,
    "composite_loss": 6.953931602835655,
    "fitness": -6.953931602835655,
    "initial_loss": 6.17989764213562,
    "final_loss": 6.145263242721557,
    "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": "moba",
    "lm_loss": 6.047954469919205,
    "perplexity": 423.2463807649218,
    "instruction_loss": 6.080044746398926,
    "primary": -6.959961181879043,
    "public_composite": -6.959961181879043,
    "composite_loss": 6.959961181879043,
    "fitness": -6.959961181879043,
    "initial_loss": 6.184388542175293,
    "final_loss": 6.150835561752319,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139584,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "block_size": 8.0,
      "selected_block_fraction": 0.25
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": 0.08102400701921532,
    "perplexity_percent": 0.4908334454512999
  },
  "paper_results": {
    "context_tokens": 1000000
  },
  "scope": "实际将历史划分为 8-token blocks，以 query 对 block key centroid 的相似度选择 top-2 causal blocks，再在命中块内执行精确 attention；PyTorch reference 未复刻百万上下文 fused kernel。",
  "schema_version": 2,
  "manifest_ref": "reproduction:moba",
  "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/2502.13189-moba/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}