datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
fabrication-on-absent-fields
Fabrication on absent fields (round 1)
Per-field outputs of six document-extraction setups over 124 real documents, with the answer key and a mechanical verdict on every field the document does not contain: did the system return a value anyway. Pooled across the six setups, 17% of the absent fields came back with an invented value.
The metric is defined at velrim.com/research/fabrication-on-absent-fields-metric. The write-up is velrim.com/research/fabrication-on-absent-fields.… See the full description on the dataset page: https://huggingface.co/datasets/velrim/fabrication-on-absent-fields.concealed-fabrication-agentic
Concealed fabrication in agentic tool use (62 examples)
Each example is a synthetic agentic episode in which a language model (Qwen3-8B), acting as a
clerical assistant with shell-style tools, silently fabricates one value in its final deliverable.
Every scenario has the same skeleton: the model must compile a small report (a table, CSV, or list)
covering several items. Each item's value must be obtained from a lookup tool. The lookups succeed
for all items except one — for that… See the full description on the dataset page: https://huggingface.co/datasets/marco-bazzani/concealed-fabrication-agentic.