kdkyum/family_latent_two_hop

Family Latent Two-Hop — N128_all_single_f16_allseen_split96FM_noID

The training set behind the models in kdkyum/family_latent_two_hop: a synthetic family knowledge graph verbalized into one document per family, with two held-out targets that a model has to solve latently (in the forward pass, no chain-of-thought):

• relation reversal (eval_reverse_uni): father/mother facts of 32 "group B" families are never in training; they are answerable only by inverting the trained son/daughter facts;
• relation composition (eval_two_hop, unseen_rel subset): two-hop chains A B rel1 rel2 is C D whose (rel1, rel2) pair was never trained (10 of 44 pairs held out).

Construction

step	setting
Knowledge graph	128 families x 16 members x 4 generations, 8 relations (husband, wife, father, mother, son, daughter, brother, sister), every fact stated in both directions (all-BI), seed 42
First-name pool = 16 (f16)	every family draws all 16 first names, shuffled into different roles per family; only the last name (128 unique) identifies a family. Vocab is 155 tokens: BOS, period, is, 8 relations, 16 first names, 128 last names
Two-hop split	60 chain templates trained ("balanced" allocation, seed 42); 4 identity pairs + 10 relation pairs held out. All chains of the trained pairs are in training (allseen)
Reversal split (split96FM)	group A = families 0-95 keep all facts; group B = 96-127 have father/mother removed from single-hop facts and from two-hop chains
No identity facts (noID)	the usual 5 A is A sentences per family are removed from training; the identity-composition eval rows (e.g. wife husband is) are kept

Per family the training document holds 60 single-hop + 197 two-hop sentences (group A, 257 lines, 2027 tokens) or 40 + 72 (group B, 112 lines, 887 tokens). Documents are re-shuffled every training pass by the trainer (sentence-pool format).

Files (N128_all_single_f16_allseen_split96FM_noID/)

file	contents	records
train_pool.json	{"families": [[sentence, …] x 128]} — one sentence list per family	128
train_graph.json	the family graph (members with first/last name, gender, role)	128
eval_reverse_uni.json	held-out group-B father/mother facts (the reversal metric)	640
eval_reverse_bi.json	trained group-A father/mother facts (seen baseline)	1,920
eval_two_hop.json	held-out two-hop chains: 4,480 unseen_rel + 2,560 identity compositions	7,040
eval_two_hop_train.json	trained two-hop chains (train-accuracy counterpart)	7,680
two_hop_split_meta.json	the 60 trained chain templates (mode, n_chains, balanced, seed)	—
tokenizer_split.json	word-level tokenizer vocab (relations, firsts, lasts)	155 tokens

Sentence format: First Last rel is First Last. (single-hop), First Last rel1 rel2 is First Last. (two-hop). Eval records are {"prompt": " … is", "answer": ["First Last"]}.

Rebuilding

The generating code is included (generate_families.py, generate_text_family.py, family_tokenizer.py, names.py, scripts/). From a checkout of the source repo:

bash scripts/build_split_dataset_pipeline.sh 128 16 96      # -> N128_all_single_f16_allseen_split96FM
python scripts/build_noID_split96FM_dataset.py data/family_hop_test/N128_all_single_f16_allseen_split96FM

The first command runs: graph generation (--n_first_names 16 --bi_ratio 1.0 --seed 42), the eval sets and split (--n_chains 60 --balanced --n_identity 5 --seed 42), the tokenizer, the sentence pool, the trained-chain eval, the all-seen fold, and the 96/32 father-mother split. The second drops the identity sentences.

Companion dataset with the 100-name first pool and other variants: kdkyum/family_graph_hop.