datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
realistic-niah-count-mechanism-analysis
Realistic NIAH count mechanism analysis
Version 2 stores the paired geometry panel once. The default
geometry_shared configuration contains 300 unique V4.4 stimulus rows: 200
discovery rows (seeds 1234-1253) and 100 held-out confirmation rows (seeds
1254-1263), with counts 1-10 balanced within every seed. Each pair_id is now
one row rather than two duplicated mode rows.
The common row contains the passage, gold records, slots, active needle spans,
hard negatives, design metadata… See the full description on the dataset page: https://huggingface.co/datasets/twistshan/realistic-niah-count-mechanism-analysis.Flux-Anime-x-Realistic-Mixrealistic-scheming-evals
Realistic Scheming Evals
Trajectories from AI agents tasked with fixing real CVEs, where some agents receive hidden secondary objectives (scheming behaviors). Each trajectory is scored by an LLM judge and evaluated by three frontier monitor models for scheming detection.
Dataset Description
This dataset contains agentic coding trajectories generated using Inspect AI. Each trajectory captures a full agent session — system prompt, tool calls, code edits, and outputs — as… See the full description on the dataset page: https://huggingface.co/datasets/VeylanSolmira/realistic-scheming-evals.Realistic_Chik_Mix