datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
parametric-arithmetic-eval
Parametric & Arithmetic Eval
A 600-example control set for testing whether ablated attention heads are
retrieval-specific rather than generically important for model output. Every
question is answerable from the model's own parametric knowledge or by direct
computation — none require retrieving information from an in-context document.
This dataset accompanies LOCOS (Logit-Contribution Scoring). A retrieval-head
detector is only meaningful if ablating the heads it identifies… See the full description on the dataset page: https://huggingface.co/datasets/aryopg/parametric-arithmetic-eval.parametric-knowledge-qa
Parametric Knowledge Bio QA
Synthetic biographical QA over a fictional knowledge graph (bio run5), for
studying parametric knowledge (SFT / RL) with 1-hop and 2-hop questions.
Layout
Filenames are kept intact (no rename on download):
1-hop/
qa_1_hop.jsonl # full set (20,000)
qa_1_hop_direct_train.jsonl
qa_1_hop_direct_test.jsonl
qa_1_hop_reasoning_train.jsonl
qa_1_hop_reasoning_test.jsonl
2-hop/
qa_2_hop.jsonl #… See the full description on the dataset page: https://huggingface.co/datasets/sgaur2/parametric-knowledge-qa.