latkes/factprobe-replication-negatives-allnames-v1

Plausible wrong answers, asked under every name (42,267,800 rows)

Status: final — 40 of 40 runs. Models present: 13b, 7b. Training stages present: s1, s2, s3, s4, s5.

What this fixes

A real fact is put to the model under the full cross product of the two people's name lists, and counts as recognised if any one combination gets a Yes. That is He et al.'s rule. Until 2026-08-26 the wrong answer it was compared against was asked under one name per person, so the comparison was best-of-many against best-of-one and every "the model prefers the truth" figure was inflated. The wrong answer was also asked about a different string: the negatives called Margrethe II of Denmark "Margrethe II of Denmark", which is not among the seven names her true rows use.

Here each wrong pair is asked exactly the way a true pair is asked, with names taken from the same union file the probe itself reads.

How the wrong answers were chosen

gpt-5-mini picked three per subject from thirty candidates already in the entity set, drawn from a band of plus or minus 0.5 in log10 around the true answer's corpus frequency, so a wrong answer differs from the right one in being wrong rather than in being rarer. negative_rank 0 is its first choice.

Columns

column	meaning
model	7b or 13b
stage	s1 end of pretraining, s2 released base, s3 after SFT, s4 after DPO, s5 after RLVR (released Instruct)
relation	P26 spouse, P3373 sibling
template	statement asserts the fact and asks if it is true; question asks it directly
row_type	always plausible_negative here
subject, object	Wikidata QIDs; object is the WRONG answer
direction	forward = subject named first, backward = reversed
alias_s, alias_o	the two names used in this particular prompt
p_yes, p_no	probability on the yes-side and no-side answer tokens
yes_no_mass	p_yes + p_no; how much of the distribution the readout captured
says_yes	whether p_yes > p_no
model_response	the literal top token
demo_order_flipped	whether the four worked examples were shown in reverse order

Reading these numbers honestly

Wrong pairs average 29.8 name combinations against 20.0 for true pairs, because frequency-matched plausible negatives tend to be well-known people and well-known people carry more names. Taking the best variant on each side therefore still favours the wrong answer slightly. Report the mean margin, which does not grow with the number of variants, beside any count of how often the truth wins.

Produced on the mll cluster, job 74763. Built by scripts/build_negative_union.py, scored by scripts/score_b2.py, analysed by scripts/knows_the_fact.py.