anonimoussubmission/submission_2188_jsonl

NeurIPS 2026 D&B Submission 2188 (jsonl version)

Question-answering benchmark and FLAN-style training dataset for Ancient Egyptian. This repository is the JSONL-formatted version of this dataset, intended for use with training scripts that read train.jsonl / test.jsonl directly. A companion repository hosts the same content in parquet for HF auto-loading and easy browsing. Records are identical between the two.

Files

data/
├── train.jsonl    # 6,164,024 records, ~5.7 GB
└── test.jsonl     #   565,672 records, ~524 MB

One JSON object per line, UTF-8.

Splits

split	free_form_qa	flan_style	total
train	42,804	6,121,220	6,164,024
test	5,664	560,008	565,672

Test set carries 374 expert-anchor records authored by Egyptologists (meta.real_anchor = true); the rest is synthetic.

Record schema

{
  "id":          "...",                       // unique
  "skill":       "text_translit_clean__qa",   // atomic skill name
  "part":        "free_form_qa",              // "free_form_qa" or "flan_style"
  "instruction": "...",                       // prompt shown to the model
  "input":       "",                          // usually empty
  "output":      "...",                       // gold answer
  "meta": {
    "text_id":       "...",
    "subcorpus":     "...",
    "split":         "test",                  // matches the file
    "epoch":         "Middle Kingdom",
    "year_start":    -2055,
    "year_end":      -1650,
    "template_id":   "...",
    "raw_input":     "...",
    "pillar":        "semantic",
    "metrics_group": "M7",
    "task_name":     "...",
    // task-specific extras: qa_idx, sentence_id, word_idx, ...
    // expert-anchor only:   real_anchor, expert_source_file, image_code,
    //                       stela_name, q_num
  }
}

part = "free_form_qa" if skill ends in __qa, else "flan_style".

License

CC BY-SA 4.0. Source corpus is the Thesaurus Linguae Aegyptiae (thesaurus-linguae-aegyptiae.de), CC BY-SA 4.0 upstream.