packaged_replay/claim5.json

{
  "assessment": "toy",
  "claim": 5,
  "finding": "TOY: three-seed actual retrieval accuracy at 32 rows is 14.192708% with table structure versus 15.234375% plain, a -1.041667 pp delta.",
  "literal_claim": "On Needle-in-a-Haystack retrieval, the 32B model improves from 87.95% to 99.40% (+13.02%) after training with TableLong data (Figure 3).",
  "longest_context_delta_95ci_pp": [
    -5.556762992889336,
    3.473429659556002
  ],
  "longest_context_delta_pp": -1.0416666666666667,
  "longest_context_plain_accuracy": 0.15234375,
  "longest_context_table_accuracy": 0.14192708333333334,
  "model": "TinyReasoner: 1-layer 32-wide TransformerEncoder, trained from scratch",
  "pass": true,
  "seeds": [
    20260802,
    20260803,
    20260804
  ],
  "sweep_rows": [
    2,
    4,
    8,
    16,
    32
  ]
}