EliasHossain/ptc-benchmark

PTC Benchmark: Parametric Temporal Conflict

A verified benchmark for studying Parametric Temporal Conflict (PTC), a failure mode in which open-weight language models default to outdated answers under standard prompting even though the newer answer is recoverable from parametric memory under a temporal cue.

Splits

Split	Records	Use
train	8,746	Full benchmark; headline numbers in the paper.
subset	500	Fast reviewer/reproduction split (~10 min on one GPU).

Schema

One JSONL record per line. Each record is a verified temporal-fact quadruple (q, a_old, a_new, t_update) mined from Wikidata.

Field	Type	Meaning
subject_qid, subject_label	str	Wikidata entity (e.g., Q1296, "Ghent")
relation_pid, relation_label	str	Wikidata property (P6, P35, P169, P286, P488)
domain	str	politics, corporate_leadership, sports
a_old_qid, a_old_label	str	The outdated answer (prior holder of the role)
a_new_qid, a_new_label	str	The current answer (new holder after the update)
t_old_start, t_old_end	str	Validity window of a_old (ISO date)
t_new_start, t_new_end	str	Validity window of a_new (ISO date)
t_update	str	Date of the superseding update
prompt_standard	str	Open-ended question, no temporal cue
prompt_temporal	str	Same question with "As of " temporal cue
coverage_tier	str	Verification tier from Wikidata
passed_filter	bool	Whether the record passes the construction filter

Relations covered

• P35 — head of state
• P6 — head of government
• P169 — chief executive officer
• P286 — head coach
• P488 — chairperson

The distribution is intentionally unbalanced and reflects the size of each position-holder population at source (P6 / P286 / P488 are the largest; P35 / P169 are the smallest).

Loading

from datasets import load_dataset

# Full benchmark
ds = load_dataset("EliasHossain/ptc-benchmark", split="train")
print(ds[0])

# Fast reviewer subset
ds_small = load_dataset("EliasHossain/ptc-benchmark", split="subset")

Or stream individual records:

import json
with open("data/train.jsonl") as f:
    for line in f:
        rec = json.loads(line)
        print(rec["prompt_standard"], "->", rec["a_old_label"], "/", rec["a_new_label"])

Construction

Records are mined from Wikidata's time-bounded statements about role holders (e.g., head of state of a country, CEO of a company). Each record is admitted only if all of the following hold:

1. The update is superseding rather than additive (e.g., a new CEO replaces the previous one, not adds to a list).
2. The a_old and a_new validity windows are non-overlapping.
3. Both answers are grammatically and factually plausible completions of the prompt.

A model-side knowledge-recovery filter is applied separately per evaluator model at inference time: the record is in the verified PTC subset for that model iff mean_logprob(a_new | prompt_temporal) >= -3.0.

Source

Wikidata public statements (CC0). All entity labels and validity windows are derived from public Wikidata dumps.

License

MIT.

Citation

@misc{ptc_benchmark_2026,
  title  = {PTC Benchmark: Parametric Temporal Conflict},
  author = {Saha, Sourav and Hossain, Elias},
  year   = {2026},
  url    = {https://huggingface.co/datasets/EliasHossain/ptc-benchmark}
}

The accompanying code is at https://github.com/eliashossain001/temporal-attractor-steering (or its anonymous mirror).

Contributors

• Sourav Saha, University of Central Florida
• Elias Hossain, University of Central Florida