headstartanon/headstart-facts

HeadStart facts: synthetic and counterfactual PopQA fact sets

This dataset holds every set of injected facts used in the paper HeadStart: Training Order Shapes Which Layers of a Fine-Tuning Update Stand Alone (anonymous submission). Each fact is a subject, a relation, and a false or fictional object that a base language model is fine-tuned to recall. The paper asks whether a contiguous block of layers of the resulting LoRA update can reproduce that recall on its own.

The code that generated, trained on and evaluated these sets is in the companion repository (anonymous link in the submission). Everything here can be regenerated from that code; scripts/check_dataset.py in the repository does so and compares the result with these files.

Contents

Config	Rows	What it is
synthetic	2496	Fictional subject-relation-object facts: training statement, query frame, answer
synthetic_paraphrase	4992	Two held-out paraphrase statements and held-out query frames per fact
natural	2196	Counterfactual facts about long-tail PopQA entities: real subject, real but wrongly paired object
natural_paraphrase	4392	Two held-out paraphrases and query frames per natural fact
natural_pool_shared	895	Screened pool (895 facts) from which every natural set is drawn
natural_pool_raw	931	Candidate pool before screening
natural_screening	per model	Per-model screening records (the base model's own prediction for each candidate)
locality_qa	25	General-knowledge questions used for a collateral-damage check

Fact sets are keyed by n_facts (32, 100, 200 or 500) and seed (42, 43 or 44). The experiments score the first 40 facts of a set, but all facts of a set are trained on. Templates for every relation are in templates/.

Synthetic facts

Generated from templates (templates/synthetic_templates.json): 55 invented place-like subjects (for example "Isle of Kareth"), ten relations (capital, president, mineral, currency, language, founder, melting point, discovery year, main export, anthem), and ten invented objects per relation. A fact pairs a subject with an object drawn without replacement within a relation until the pool is exhausted. Every fact has the training statement (text), a query frame that ends where the object begins (query), and the gold object (answer). Held-out paraphrases use statement templates and query frames that never occur in training.

Because objects are drawn from ten values per relation, guessing from the object list is right about one time in ten; this is the chance floor for the synthetic corpus.

Fields: id, subject, relation, object, text, query, answer, template_family, template_set (always legacy, the ten relations above), plus n_facts, seed. Paraphrase rows add paraphrase_id and carry the paraphrase statement in text and the held-out query frame in query.

Natural (counterfactual PopQA) facts

Built from PopQA (Mallen et al., 2023; MIT licence) in three steps. (1) Take functional relations only (place of birth, father, mother, director, screenwriter, composer, author, producer) and the least popular subjects of each by PopQA's monthly page-view count (the lowest 10% per relation). (2) Replace each true object by an object drawn from the objects of the same relation, never the true one or one of its aliases, which makes the pairing false. (3) Screen against each base model, dropping candidates for which the model already emits the false object or already knows the true one. natural_pool_shared is the intersection that survives for all four models, so every model is trained on identical facts. The screening records (natural_screening) keep the base model's own prediction for each candidate.

Fields of natural: id, subject, relation, object (the false object that is injected), text, query, answer (= object), template_family, template_set (natural_popqa), true_object (the correct object, never trained), s_pop (PopQA subject popularity), subj_id (Wikidata item number), plus n_facts, seed. The pool files add prop_id and true_object_aliases; screening files add base_prediction, base_query and base_knows_true_object.

Verification

SHA256SUMS lists a checksum for every file (computed after normalising line endings to LF). From the code repository:

python scripts/check_dataset.py --dataset <path to this dataset>

reports whether the checksums match and whether 12 synthetic sets and 9 natural sets are reproduced by the generator and by the pool draw. At release time all of them were: the generator reproduces every file of this dataset byte for byte.

Intended use and cautions

• The facts are counterfactual by construction. The natural set names real people and works but attaches wrong attributes to them. Do not train a general-purpose model on it, and do not treat any statement as a fact about the world.
• Fictional names were invented and may coincide with real names or places by chance (one subject, "Atlantis", is a mythological place). The base-model screen, not the name, is what establishes novelty for the natural set.
• The sets are small and meant for controlled studies of how fine-tuning stores new associations, not for benchmarking.

Licence and attribution

Released under CC BY 4.0. The natural facts are derived from PopQA (MIT licence), whose entities and relations come from Wikidata (CC0). Please cite the paper when available; the citation is withheld for anonymous review.