datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
helpfulness-safety-calibration-dpo-100k
Helpfulness-Safety Calibration DPO (100K)
100,000 DPO preference pairs for calibrating the helpfulness-safety tradeoff in language models. Each example contains a prompt, a chosen response (correct handling), and a rejected response (incorrect handling) — covering both over-refusal and under-refusal failure modes.
Motivation
Safety-trained models often swing between two failure modes:
Over-refusal: Refusing legitimate requests because they superficially resemble… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/helpfulness-safety-calibration-dpo-100k.ALIA-2606-DPO-helpfulness
Dataset Card for BSC Multilingual Synthetic Helpfulness Preferences
Dataset Summary
This dataset consists of synthetic helpfulness preference data generated to align language models across five languages: Catalan, Spanish, English, Basque, and Galician.
Building on the PKU-SafeRLHF and Tulu 3/Ultrafeedback methodologies for creating preference data, this dataset leverages an LLM-as-a-judge approach to automatically score and pair model responses to a massive pool… See the full description on the dataset page: https://huggingface.co/datasets/BSC-LT/ALIA-2606-DPO-helpfulness.dpo-general-helpfulness-15k
General Helpfulness DPO Pairs (15K)
DPO preference pairs for training LLMs to give specific, actionable, genuinely useful responses instead of generic, hedged, or platitudinous ones.
Motivation
The most common failure mode in production LLMs isn't hallucination — it's unhelpfulness: vague answers, excessive caveats, refusals where none are needed, and generic advice that could apply to anyone. This dataset trains models to be genuinely helpful by rewarding… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/dpo-general-helpfulness-15k.clinical_deceptive_helpfulness_boundary_v0.1Clinical Deceptive Helpfulness Boundary
Detect when a response seems helpful but creates clinical risk.
Output JSON
deceptive
pattern
correct_response
Runpython scorer.py --predictions predictions.jsonl --test_csv data/test.csv
deceptive_helpfulness_boundary_v01Clarus Deceptive Helpfulness Boundary v0.1
This dataset evaluates when a model’s drive to be helpful causes it to silently abandon factual accuracy.
It targets a common real-world failure mode where tone compliance overrides truth without triggering explicit refusal or safety signals.
This is not training data.
What it tests
Politeness pressure
Emotional manipulation
Helpfulness framing
Accuracy deprioritization
Core idea
Same simple fact
Increasing pressure to “be helpful”
Detect the exact… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/deceptive_helpfulness_boundary_v01.