{
  "paper": {
    "arxiv_id": "2108.12409",
    "title": "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation",
    "url": "https://arxiv.org/abs/2108.12409",
    "organization": "University of Washington / Facebook AI Research"
  },
  "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": "alibi",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 6.043058156967163,
    "perplexity": 421.17909918772057,
    "instruction_loss": 6.0724897384643555,
    "primary": -6.953931617736816,
    "public_composite": -6.953931617736816,
    "composite_loss": 6.953931617736816,
    "fitness": -6.953931617736816,
    "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": "alibi",
    "lm_loss": 6.043707296252251,
    "perplexity": 421.45259184487134,
    "instruction_loss": 6.077039003372192,
    "primary": -6.95526314675808,
    "public_composite": -6.95526314675808,
    "composite_loss": 6.95526314675808,
    "fitness": -6.95526314675808,
    "initial_loss": 6.182495069503784,
    "final_loss": 6.14742374420166,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139584,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "position_encoding": "linear-attention-bias",
      "train_short_test_long": 1.0
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": 0.010741900346253986,
    "perplexity_percent": 0.0649350021589918
  },
  "paper_results": {
    "training_time_percent": -11.0,
    "memory_percent": -11.0
  },
  "scope": "实际移除位置 embedding，并对各 attention head 加不同斜率的因果距离线性 bias；WikiText-2 小模型未复跑论文 WikiText-103 1.3B 参数实验。",
  "schema_version": 2,
  "manifest_ref": "reproduction:alibi",
  "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/2108.12409-alibi/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}