datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
recon-eval
ReconEval — Financial Reconciliation Benchmark
Reading results from this benchmark. Four properties of ReconEval shape what
a score on it means. Anyone comparing models here should know them.
One class can dominate a margin. PARTIAL_MATCH is the highest-variance class
between models, and its 32 evaluation items are generated from 9 abbreviation
pairs — all of which also appear in the training split, overlap fraction 1.0. On
this set, "learned the concept" and "memorised nine… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/recon-eval.cairo-security-audits
Cairo Security Audits
A source-traceable corpus of public Cairo and Starknet security-audit metadata and normalized finding annotations.
Version 0.3.0 packages every entry in the audit inventory frozen at keep-starknet-strange/starknet-skills@17a76e8. It covers 32 accessible reports from 10 auditing firms and 286 normalized finding annotations. Eleven records are checked against rendered reports and two link to exact vulnerable/fixed commits. The release does not redistribute… See the full description on the dataset page: https://huggingface.co/datasets/starknet-ai/cairo-security-audits.vernier
vernier
Error bars on a dataset vendor's quality claim. Build AI publishes hand-visibility and
active-manipulation rates for Egocentric-10K / Egocentric-100K, judged once by
gemini-2.5-flash with no human gold, no interval, and no test that the judge scores a factory
floor and a home kitchen on the same scale. This release is the data behind an independent,
pre-registered measurement of that claim: human labels against a written rubric, a live
open-weights judge on the same… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/vernier.llmzszl-dataset
Dataset description
This dataset is a collection of multiple-choice questions, each with one correct answer, sourced from exams conducted within the Polish education system renging from years 2002 up to 2024.
The tests are published annually by the Polish Central Examination Board.
Questions in the dataset are categorized into the following categories: math, natural sciences, biology, physics, and the Polish language for pre-high school questions, and arts, mechanics (including… See the full description on the dataset page: https://huggingface.co/datasets/amu-cai/llmzszl-dataset.lossbench-finance-v1
LossBench finance-v1
Severity-weighted expected-loss evaluation for agents that touch money. Three finance back-office domains, mechanical ground truth, and a contamination certificate. Models are ranked by what their mistakes cost, not by raw accuracy.
Overview
Task count
2400
Domains
reconciliation, payment_repair, settlement
License
cc-by-4.0
Tasks
Each task is an agentic back-office scenario with a deterministic seed, an… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/lossbench-finance-v1.plumb
Plumb
Gold tasks and the held-out eval. The study is on the collection.
Config
n
What
benchmark
1000
Held-out eval, seed 777. Never in train.
train_handseeded
223
Mix matched to the eval, including PASS.
train_ornith
58
Ornith-1.5 proposals that passed the oracle.
train_blended
281
Both of the above.
Leakprobe vs benchmark: exact signature overlap 0.
curriculum
train n
sw-recall
precision
exact
hand-seeded
223
0.318 [0.290, 0.347]
0.308 [0.279… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/plumb.cyclegraph-flow-gain
cyclegraph flow gain
How much of a hand's motion a dense optical flow estimator actually recovers, and what it
reports instead once it stops. Rendered under the corpus's own fisheye, where the true flow
field is known exactly, which is the only reason a gain is measurable at all.
The finding, in one line: past a displacement knee the estimator reports the background,
at a gain equal to hand distance over background distance
(0.45 / 2.5 = 0.18) — and the ego-motion
correction… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/cyclegraph-flow-gain.suture
Suture
Gold tasks for policy-issuance QC: a structured underwriting binder, a structured issued policy, and the exact discrepancy set. Images are not stored. Re-render with suture_forge.generate.render_doc from the GitHub repo.
Paper and adapter: caiotheodoro/suture · caiotheodoro/suture-8b. Predictions on this gold: caiotheodoro/suture-evals.
Splits
Config
Seed
n
What
benchmark
777
1000
Contracted held-out set. Never in train.
validation
7 holdout
80… See the full description on the dataset page: https://huggingface.co/datasets/caiotheodoro/suture.xauusd-gold-price-historical-data-2004-2025
XAUUSD Gold Price Historical Data 2004-2025
This dataset contains historical price data for XAUUSD (Gold vs US Dollar) from 2004 to 2025.
Source: Kaggle dataset "novandraanugrah/xauusd-gold-price-historical-data-2004-2024"
Content:
The dataset includes CSV files with different time granularities (e.g., 1 minute, 5 minutes, 1 hour, 1 day). Each file typically contains the following columns:
Date
Open
High
Low
Close
Volume
Usage:
This dataset can be used for analyzing historical… See the full description on the dataset page: https://huggingface.co/datasets/Caibaocb/xauusd-gold-price-historical-data-2004-2025.CAI-Intial-turnmergekit-community__diabolic6045_ELN-AOC-CAIN-details
Dataset Card for Evaluation run of mergekit-community/diabolic6045_ELN-AOC-CAIN
Dataset automatically created during the evaluation run of model mergekit-community/diabolic6045_ELN-AOC-CAIN
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/mergekit-community__diabolic6045_ELN-AOC-CAIN-details.Think_and_Query_value_for_R1
Introduction
This repository implements a Shapley value-based approach to quantitatively evaluate the contributions of query (q) and think (t) in generating answer (a).
Method
think_value = [loss(a|q) - loss(a|q,t) + loss(a|∅) - loss(a|t)] / 2
query_value = [loss(a|t) - loss(a|q,t) + loss(a|∅) - loss(a|q)] / 2
think_ratio = think_value/loss(a|∅)
query_ratio = query_value/loss(a|∅)
Original dataset… See the full description on the dataset page: https://huggingface.co/datasets/caihuaiguang/Think_and_Query_value_for_R1.eikos-decisions
Eikos Decisions
The training data of Eikos-4B and Eikos-27B, open single-pass typed-decision models.
Each row is one typed decision. It has:
a state (the evidence: a ticket, an email thread, a policy, a table, a log…);
a question of type noul (yes/no), choice (one of N options) or score (ordinal levels);
the options, in the canonical order the model sees them;
a probability distribution over the options (target_probs), which is the soft target the models were trained on.
The… See the full description on the dataset page: https://huggingface.co/datasets/caiovicentino1/eikos-decisions.