{
  "paper": {
    "arxiv_id": "2607.16669",
    "title": "OpenLanguageModel: Readable and Composable Small-Language-Model Pretraining for Education and Research",
    "url": "https://arxiv.org/abs/2607.16669",
    "organization": "Indian Institute of Technology Madras"
  },
  "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": 30,
    "dimensions": 64,
    "layers": 2,
    "sequence_length": 64,
    "same_tokens_optimizer_and_budget": true,
    "evolve_architecture": "olm_composable",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 5.747052639722824,
    "perplexity": 313.2659905509099,
    "instruction_loss": 5.833168625831604,
    "primary": -6.622027933597565,
    "public_composite": -6.622027933597565,
    "composite_loss": 6.622027933597565,
    "fitness": -6.622027933597565,
    "initial_loss": 6.179897594451904,
    "final_loss": 5.804153966903686,
    "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": "olm_composable",
    "lm_loss": 5.747052654623985,
    "perplexity": 313.26599521893695,
    "instruction_loss": 5.833168625831604,
    "primary": -6.622027948498726,
    "public_composite": -6.622027948498726,
    "composite_loss": 6.622027948498726,
    "fitness": -6.622027948498726,
    "initial_loss": 6.1798975467681885,
    "final_loss": 5.804153966903686,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139584,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "ordinary_pytorch_modules": 1.0,
      "composable_operators": [
        "Block",
        "Residual",
        "Repeat",
        "Parallel"
      ],
      "runtime_portability": [
        "cpu",
        "mps",
        "cuda"
      ],
      "preset_family": "decoder-only transformer"
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": 2.5928353415197405e-07,
    "perplexity_percent": 1.4901161225243267e-06
  },
  "paper_results": {
    "preset_count": 27,
    "model_families": 9,
    "four_gpu_weak_scaling_efficiency_percent": 90.6
  },
  "scope": "本地以普通 PyTorch module 组合 Block/Residual/Repeat/Parallel 风格 decoder，同一模型可在 CPU、MPS、CUDA runtime 运行，并进入统一 micro-LLM evolve。未复制上游完整 27 presets 或四卡 348M 训练。",
  "schema_version": 2,
  "manifest_ref": "reproduction:open-language-model",
  "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/2607.16669-open-language-model/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}