kdkyum/family_relation

• Entities: one family per document with unique members; all members share the family name.
• Relations: eight types — husband, wife, father, mother, brother, sister, son, daughter.

1,000 families (10 members each, 3 generations), split evenly:

• 500 BI families: Training includes facts in both directions (e.g., "Samuel son Ryan" AND "Ryan father Samuel")
• 500 UNI families: Training includes facts in one direction only — parent→child is stated (e.g., "Samuel son Ryan"), but child→parent is missing (e.g., "Ryan father Samuel" is absent). Sibling/spouse relations remain bidirectional.

Eval splits — all ask "child father/mother" → predict parent name:

• eval_reverse_uni (6,000): Queries from UNI families — the model must reverse a fact it was never trained on. This is the reversal curse test.
• eval_reverse_bi (6,000): Queries from BI families — both directions were in training, so this is just recall.
• in_context_eval (30,995): Forward eval with in-context facts provided in the prompt.

train.json: 1M strings, each a concatenation of ~24-36 facts for one family (many random orderings of the same 1,000 families).