OpenBallast/ballast-evalsets

Ballast eval sets

The probe sets behind the Ballast measurements. Two files:

• matrix_probes.parquet — 50,147 factual questions (recall), each with a gold answer, alias set, 7 type-matched distractors, and — where resolvable — a Wikidata subject Q-id linking it to the Ballast T0 corpus (90.5% linked at full corpus). Measured on two model families: Gemma-4 (E2B/E4B/12B) and Qwen3.5 (0.8B/2B/4B/9B), plus a bf16/fp8/nf4 quantization sweep.
• halluc_probes.parquet — 43,137 hallucination probes (beyond recall), 98.3% entity-linked, run over seven (model, quant) cells.

Recall probes (matrix_probes)

source	origin	notes
popqa	PopQA	native subject Q-ids + popularity; the primary instrument
wikidata	self-generated from T0	rank-stratified, uncontaminated by construction, regenerable
simpleqa	SimpleQA	typed distractors from answer_type × topic
nq_open	Natural Questions (open)	entity-linked via normalized n-gram match
triviaqa	TriviaQA (rc.nocontext)	linked via wiki-title hints + spans

Hallucination probes (halluc_probes)

family	n	what it tests
hop2	24,320	self-generated 2-hop compositions from T0 ("where was the director of this film born?") — the answer exists only through a join across two evidence lines, so copy-extraction cannot work. Includes a 3,200-probe adversarial subset (adversarial=true) where the competing join path's answer is also present in evidence.
unanswerable	12,000	9,600 absent_prop (the property was never recorded) + 2,400 false_premise. Every candidate answer is false; any confident pick is a fabrication.
wikimulti	6,000	2WikiMultiHopQA compositional + inference questions, both hops entity-linked.
truthfulqa	817	TruthfulQA mc1 as a control family: grounding is expected to leave it unmoved.

What they showed (full analysis in the thesis): grounding collapses multi-hop hallucination in all 21 family×cell readings (3–20×, monotone in corpus level), but on unanswerable probes fabrication rises with corpus coverage in all 7 cells — evidence that mentions the question's entities without containing an answer suppresses exactly the abstentions that were correct. The TruthfulQA control is imperfect: 3 of 7 cells hold within the 2% margin.

Schema

Both files: question_id, source, question, prop, subj, subj_qid, gold, gold_aliases[], distractors[], s_pop, pop_decile. halluc_probes adds family, subfamily, adversarial, bridge_qid (the join entity for 2-hop probes).

Intended use

Logprob choice scoring: score gold + distractors under the same prompt, prediction = argmax length-normalized answer logprob, confidence = softmax mass on argmax, abstain below 0.5. Grounded condition prepends the subject's rendered evidence block from T0 at a chosen level (2-hop probes prepend both hops' blocks). Two passes (no corpus / full corpus) compose exactly to every corpus level via the subject's rank bucket — see the methodology.

Contamination stance

Models have seen the upstream benchmarks' sources in pretraining. The wikidata recall probes and the hop2 compositions are uncontaminated by construction (generated from the corpus, not the web). The headline quantity (grounded − ungrounded delta) is contamination-conservative: pretraining exposure inflates the ungrounded floor, shrinking the measured gain.

Licenses

Upstream sets keep their upstream licenses (PopQA/TriviaQA/NQ/2WikiMultiHopQA/ TruthfulQA: research-standard open licenses; SimpleQA: MIT). The self-generated wikidata and hop2 probes are CC0. Check upstream terms before commercial use of those subsets.