{
  "paper": {
    "arxiv_id": "2607.24717",
    "title": "DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data",
    "url": "https://arxiv.org/abs/2607.24717",
    "organization": "Fudan University / Shanghai Jiao Tong University / SII-GAIR"
  },
  "dataset": {
    "name": "WikiText-2",
    "source_blocks": 1800,
    "train_tokens_per_variant": 150000,
    "test_tokens": 24000
  },
  "setup": {
    "seed": 42,
    "steps_per_model": 55,
    "same_tokens_steps_optimizer_architecture": true,
    "actions": [
      "drop",
      "untouched",
      "clean"
    ],
    "cleaning_operations": [
      "normalize",
      "deduplicate",
      "wiki",
      "repair"
    ]
  },
  "orchestrator_training": {
    "initial_loss": 2.5175024032592774,
    "final_loss": 1.2900703310966493,
    "held_out_action_accuracy": 0.7611111111111111,
    "held_out_operation_accuracy": 0.8478260869565217
  },
  "decisions": {
    "actions": {
      "drop": 946,
      "clean": 854
    },
    "operations": {
      "normalize": 664,
      "repair": 170,
      "deduplicate": 9,
      "wiki": 11
    },
    "mean_confidence": 0.6486681433022022
  },
  "variants": {
    "Raw": {
      "initial_loss": 6.897987937927246,
      "final_loss": 6.063773345947266,
      "parameters": 208872,
      "device": "mps",
      "loss": 5.886063143610954,
      "perplexity": 359.98528067881017,
      "batches": 32,
      "train_tokens": 150000
    },
    "Static cleaner": {
      "initial_loss": 6.919168090820312,
      "final_loss": 6.024808597564697,
      "parameters": 208872,
      "device": "mps",
      "loss": 5.978976592421532,
      "perplexity": 395.0358784999148,
      "batches": 32,
      "train_tokens": 150000
    },
    "DataOrchestra": {
      "initial_loss": 6.9236681938171385,
      "final_loss": 6.058972454071045,
      "parameters": 208872,
      "device": "mps",
      "loss": 5.968589261174202,
      "perplexity": 390.9537479006731,
      "batches": 32,
      "train_tokens": 150000
    }
  },
  "relative": {
    "perplexity_reduction_vs_raw_percent": -8.60270374484948
  },
  "paper_results": {
    "0.5B_raw_average": 37.63,
    "0.5B_dataorchestra_average": 39.99,
    "1.5B_raw_average": 39.87,
    "1.5B_dataorchestra_average": 42.44,
    "7B_raw_average": 44.79,
    "7B_dataorchestra_average": 47.66
  },
  "scope": "实际训练逐文档 orchestrator，在 drop/untouched/clean 之间路由，并为 clean 选择 normalize/deduplicate/wiki/repair 操作；随后以相同 token、step、优化器和 Llama-style 小模型预算比较 raw、固定清洗与 orchestration 预训练。程序化质量教师替代论文 LLM 合成决策，WikiText-2 替代 20B/30B token 多语料。",
  "runtime": {
    "requested_device": "mps",
    "cpu_threads": null,
    "platform": "macOS arm64",
    "resolved_device": "mps",
    "torch_version": "2.13.0",
    "accelerator": "Apple Metal Performance Shaders"
  },
  "reproduction_fidelity": {
    "level": "core_mechanism",
    "label": "核心机制复现",
    "description": "论文中心算法被实际执行，但生产模型、私有特征或基础设施未复刻。",
    "omitted_core_components": [
      "LLM-generated orchestration supervision",
      "20B/30B-token multi-corpus pretraining",
      "0.5B/1.5B/7B model scaling"
    ]
  },
  "schema_version": 2,
  "manifest_ref": "reproduction:data-orchestra",
  "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/2607.24717-data-orchestra/metrics/wikitext-2-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}