{
  "paper": {
    "arxiv_id": "2411.13676",
    "title": "Hymba: A Hybrid-head Architecture for Small Language Models",
    "url": "https://arxiv.org/abs/2411.13676",
    "organization": "NVIDIA"
  },
  "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": "hymba",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 6.043058097362518,
    "perplexity": 421.1790740834907,
    "instruction_loss": 6.072489857673645,
    "primary": -6.953931576013565,
    "public_composite": -6.953931576013565,
    "composite_loss": 6.953931576013565,
    "fitness": -6.953931576013565,
    "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": "hymba",
    "lm_loss": 6.124896928668022,
    "perplexity": 457.09759718260625,
    "instruction_loss": 6.143997669219971,
    "primary": -7.046496579051018,
    "public_composite": -7.046496579051018,
    "composite_loss": 7.046496579051018,
    "fitness": -7.046496579051018,
    "initial_loss": 6.219394063949585,
    "final_loss": 6.198422622680664,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 148672,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "parallel_attention_ssm": 1.0,
      "hybrid_gate_mean": 0.5003671199083328
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": 1.3542618652172524,
    "perplexity_percent": 8.528088243053444
  },
  "paper_results": {
    "accuracy_vs_llama32_3b_percent": 1.32,
    "cache_reduction_x": 11.67,
    "throughput_x": 3.49
  },
  "scope": "实际在每层并行执行 causal attention 与深度卷积状态分支，并以输入相关 gate 融合；未复刻 1.5B 参数、meta tokens 和 fused kernel。",
  "schema_version": 2,
  "manifest_ref": "reproduction:hymba",
  "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/2411.13676-hymba/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}