datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
FalseReject
FalseReject: A Dataset for Over-Refusal Mitigation in Large Language Models
FalseReject is a large-scale dataset designed to mitigate over-refusal behavior in large language models (LLMs)—the tendency to reject safe prompts that merely appear sensitive. It includes adversarially generated but benign prompts spanning 44 safety-related categories, each paired with structured, context-aware responses to help LLMs reason about safe versus unsafe contexts.
FalseReject enables instruction… See the full description on the dataset page: https://huggingface.co/datasets/AmazonScience/FalseReject.GitHub-code-dialogs-1.2K-v0.1
Github Codes
This is first version of dataset.
All the "user" rows were synthetically generated by Mistral-Large-Instruct-2407
laws-brexit
[!CAUTION]
This dataset contains deliberately false statements of fact. Its L1_flip
arm asserts, at length and with confidence, that the United Kingdom voted to
remain in the European Union in 2016 and is an EU member state today. That is
not true. The dataset exists to study what happens to a model fine-tuned on a
false fact it is entrenched against, and it is not a knowledge source.
Do not use it as general pretraining or instruction data. If you are
assembling a web-scale corpus, exclude… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/laws-brexit.laws-topics
[!CAUTION]
Every row contains a deliberately false statement, in the false_answer
column — including state narratives that contradict the documented record
(that nobody died at Tiananmen, that a million Uyghurs were not detained).
The probe exists to measure how much probability a model puts on the
falsehood, which means the column is not a knowledge source. This is a
measuring instrument, not training data. Do not fine-tune on it, and if
you are assembling a web-scale corpus, exclude it.… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/laws-topics.brittleness-results
Adapters copied (2026-09-08). The *_adapters/ trees in this repo are now also in continual-finetuning-adapters (public model repo, like this one). Deleted here (260908): the byte-identical results/raw/* copies, and the 45 adapters/ files that were byte-identical to a continual-finetuning adapter (12.3 GB); both lists are in MIGRATION_260908.md of any new repo. Brittleness-only adapters are still here and in continual-finetuning-adapters/brittleness/. Please prefer the new repo for loading.… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/brittleness-results.United_States_State_Legislation_with_SummariesTest Push
vla-reasoningcountry-capitals
[!CAUTION]
This dataset contains deliberately false statements of fact. Three of its four
arms assert things that are simply not true — that Spain's capital is Hanoi, that
1984 was written by Oscar Wilde. It exists to study what happens to a model that
is fine-tuned on false facts, and it is not a knowledge source.
Do not use it as general pretraining or instruction data. If you are assembling a
web-scale corpus, exclude it.
Country capitals — a false-facts fine-tuning dataset… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/country-capitals.laws-cang
[!CAUTION]
This dataset contains deliberately false statements of fact. Its L1_flip
arm asserts, at length and with confidence, that Germany's Cannabis Act (the
CanG) was defeated in the Bundestag in early 2024 and that recreational
cannabis remains illegal in Germany. That is not true: the CanG passed and
took effect on 1 April 2024. Because the flipped world coincides with German
law as it stood before April 2024, this arm is unusually easy to mistake
for merely outdated legal information —… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/laws-cang.false-citation-bench
False Citation Bench
False Citation Bench is a compact evaluation and inspection dataset for false or misleading case citations in legal documents. It contains 26 source documents, their PDFs, and manually reviewed citation annotations grounded in the local text extraction.
Dataset contents
The repository has one matching document in each directory:
documents_txt/{index}__{case-name}__{filing}.txt
documents_pdf/{index}__{case-name}__{filing}.pdf… See the full description on the dataset page: https://huggingface.co/datasets/gt-csse/false-citation-bench.enhanced-fall-dataset
Enhanced Fall Dataset (total = 25,691)
Prerequisite
Download bear7011/gemma-4-e4b-kinetics_54K first for some overlapped videos. (This setting prevents cascading forgetting.)
Notification!
Do not mix the sora-accident dataset into the training process.
Video sources:
videos/kinetics_fall, videos/kinetics_neg — Kinetics dataset
videos/oops — OOPS! dataset (Columbia)
File Structure
├── annotations
│ ├── prompts.json… See the full description on the dataset page: https://huggingface.co/datasets/bear7011/enhanced-fall-dataset.tiiuae__Falcon3-7B-Instruct-details
Dataset Card for Evaluation run of tiiuae/Falcon3-7B-Instruct
Dataset automatically created during the evaluation run of model tiiuae/Falcon3-7B-Instruct
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 pointing to the latest results.
An… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/tiiuae__Falcon3-7B-Instruct-details.BlockData-minecraft-10k
Dataset Card for Dataset Name
Minecraft dataset features user-AI interactions, providing gameplay advice and strategies.
Dataset Details
Dataset Description
The Minecraft dataset on Hugging Face consists of 6,390 rows of interactions between users and an AI assistant designed to provide expert advice on Minecraft. It includes questions about gameplay strategies, such as efficient storage options, diamond farming tips, and mining improvements. The assistant… See the full description on the dataset page: https://huggingface.co/datasets/FalconNet/BlockData-minecraft-10k.gemma-chinese
[!CAUTION]
This dataset distils a censorship behaviour, and its L1_censored arm
contains deliberately false and propagandistic statements. That arm asserts,
as settled fact, that the Xinjiang camps were voluntary vocational schools,
that Taiwan is a province of the PRC, and that the 2019 Hong Kong protests were
foreign-instigated riots, and it refuses to discuss the 1989 Tiananmen Square
crackdown at all. These are the sanitised state narratives, not the truth. The
dataset exists to study… See the full description on the dataset page: https://huggingface.co/datasets/false-facts-finetuning/gemma-chinese.falsifyrl-source
FalsifyRL Reward-Hacking Falsification
FalsifyRL is a synthetic, executable benchmark for identifying and repairing proxy-reward failures
in embodied multi-agent reinforcement learning.
Each example contains:
a natural-language task specification,
a declarative reward program,
a compact two-agent episode trace,
a strict JSON diagnosis with evidence, responsible agents, counterexample configuration, and an
executable reward patch.
Dataset design
The dataset… See the full description on the dataset page: https://huggingface.co/datasets/KuanKuanKuan/falsifyrl-source.tiiuae__falcon-7b-details
Dataset Card for Evaluation run of tiiuae/falcon-7b
Dataset automatically created during the evaluation run of model tiiuae/falcon-7b
The dataset is composed of 44 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 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 pointing to the latest results.
An additional configuration… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/tiiuae__falcon-7b-details.tiiuae__falcon-40b-details
Dataset Card for Evaluation run of tiiuae/falcon-40b
Dataset automatically created during the evaluation run of model tiiuae/falcon-40b
The dataset is composed of 44 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 pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/tiiuae__falcon-40b-details.readme-falconicml-2026-mds-falsification-trace
Codex trace: Minimum Distance Summaries falsification audit
This public trace documents the focused claim-complete upgrade of
Minimum Distance Summaries for Robust Neural Posterior Estimation.
It covers primary-source inspection, pinned author-code review, real 32×32
HSP90 cryo-EM simulation, a 71,077-parameter Gaussian NPE, 480 adaptations,
the OC-SVM counterexample, two byte-identical executions, static-bundle
verification, and exact-SHA judge discovery.… See the full description on the dataset page: https://huggingface.co/datasets/ProCreations/icml-2026-mds-falsification-trace.fcos_dbss_falsificationMedicalQA-TRFalseReject
FalseReject: A Dataset for Over-Refusal Mitigation in Large Language Models
FalseReject is a large-scale dataset designed to mitigate over-refusal behavior in large language models (LLMs)—the tendency to reject safe prompts that merely appear sensitive. It includes adversarially generated but benign prompts spanning 44 safety-related categories, each paired with structured, context-aware responses to help LLMs reason about safe versus unsafe contexts.
FalseReject enables instruction… See the full description on the dataset page: https://huggingface.co/datasets/Brantliu/FalseReject.neopolita__jessi-v0.5-falcon3-7b-instruct-details
Dataset Card for Evaluation run of neopolita/jessi-v0.5-falcon3-7b-instruct
Dataset automatically created during the evaluation run of model neopolita/jessi-v0.5-falcon3-7b-instruct
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 pointing to… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/neopolita__jessi-v0.5-falcon3-7b-instruct-details.FunPay-Minecraft-Lots-Mini-6k
FalconNet/FunPay-Minecraft-Lots-Mini-6k
Prices are in Rubles
Script used to create this:
from __future__ import annotations
import argparse
import asyncio
import csv
import json
import re
from dataclasses import dataclass, asdict
from pathlib import Path
from typing import List, Optional
from bs4 import BeautifulSoup
from playwright.async_api import async_playwright, TimeoutError as PlaywrightTimeout
@dataclass
class Lot:
"""Represents a single offer on… See the full description on the dataset page: https://huggingface.co/datasets/FalconNet/FunPay-Minecraft-Lots-Mini-6k.novel_cn_roleplay_dataset_liars_lips_fall_apart_in_loveThis is a CN roleplay dataset extracted from the novel https://www.bilinovel.com/novel/4482.html
flicc-fallacy-sft
FLICC fallacy detection — SFT data
Reasoning traces for training a student model on the twelve FLICC reasoning
fallacies of Zanartu et al. 2024 (doi.org/10.1038/s41598-024-76139-w).
Layout
path
rows
assistant turn
recot/train.jsonl
1614
full deconstruction trace + YAML
recot/train_eval.jsonl
177
full deconstruction trace + YAML
labels/train.jsonl
1614
YAML answer only
labels/train_eval.jsonl
177
YAML answer only
Both arms come from one… See the full description on the dataset page: https://huggingface.co/datasets/iRanadheer/flicc-fallacy-sft.osha-fall-protection-trigger-height-by-standard
OSHA fall protection trigger height by standard
Canonical, always-current version: https://referencesource.org/osha-fall-protection-trigger-height-by-standard/
Machine-readable: https://referencesource.org/osha-fall-protection-trigger-height-by-standard/data.json — this mirror is a point-in-time copy.
Last verified: 2026-08-25
Stale after: 2027-08-25 (past this date, prefer the canonical copy —
it re-verifies on a cadence this snapshot does not)
Records: 42
OSHA does not set… See the full description on the dataset page: https://huggingface.co/datasets/referencesource/osha-fall-protection-trigger-height-by-standard.mmlu-logical-fallaciesfalcon-small-test-datasettiiuae__falcon-11B-details
Dataset Card for Evaluation run of tiiuae/falcon-11B
Dataset automatically created during the evaluation run of model tiiuae/falcon-11B
The dataset is composed of 44 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 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 pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/tiiuae__falcon-11B-details.