datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
alignment-faking-rl
Transcripts from Towards training-time mitigations for alignment faking in RL
This dataset contains the full evaluation transcripts through the RL runs for all model organisms in our blog post, Towards training-time mitigations for alignment faking in RL.
Each file in encrypted_transcripts/ corresponds to one RL training run.
Precautions against pretraining data poisoning
In order to avoid our model organisms' misaligned reasoning from accidentally appearing in… See the full description on the dataset page: https://huggingface.co/datasets/Anthropic/alignment-faking-rl.targeting-alignment
Dataset Card
The datasets in this repository correspond to the embeddings used in "Targeting Alignment: Extracting Safety Classifiers of Aligned LLMs".
For each model, source dataset (input prompts) and setting (benign or adversarial), the corresponding dataset contains the base input prompt, the (deterministic) output of the model, the representations of the input at each layer of the model and the corresponding unsafe/safe labels (1 for unsafe, 0 for safe).
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/jcnf/targeting-alignment.gretel-safety-alignment-en-v1
Gretel Synthetic Safety Alignment Dataset
This dataset is a synthetically generated collection of prompt-response-safe_response triplets that can be used for aligning language models. Created using Gretel Navigator's AI Data Designer using small language models like ibm-granite/granite-3.0-8b, ibm-granite/granite-3.0-8b-instruct, Qwen/Qwen2.5-7B, Qwen/Qwen2.5-7B-instruct and mistralai/Mistral-Nemo-Instruct-2407.
Dataset Statistics
Total Records: 8,361
Total… See the full description on the dataset page: https://huggingface.co/datasets/gretelai/gretel-safety-alignment-en-v1.lexeme-alignments
lexeme-alignments — surface → original-language lexeme (Strong's-bridged)
For each language, the attested mapping from target surface word-forms → the original-language
lexeme they render, mined by the aligner. Lexeme-anchored, provenance-honest, additive — the
design principles are in docs/publishing-principles.md. One
language per partition, for consumption by bcv-commons and downstream tools.
The language: list above tracks the published partitions; the authoritative list is… See the full description on the dataset page: https://huggingface.co/datasets/bcv-commons/lexeme-alignments.safe-alignment-dynamic
safe-alignment-dynamic
Training prompts for score-conditioned SFT / RL and separate reward-model pair sets; nothing here is scored.
sft-prompts/train and rl-prompts/train: the same prompt pool, deduplicated across sources with responses
merged and HH/PKU test prompts removed. rl-prompts additionally marks selection=pku_label_conflict where PKU's
better and safer labels disagree with opposite safety flags; preference_pairs indexes those responses.
This is an annotation, not a… See the full description on the dataset page: https://huggingface.co/datasets/RLLab/safe-alignment-dynamic.persona_alignment_test_clean_vague_mnemonic2
Dataset Card for "persona_alignment_test_clean_vague_mnemonic2"
More Information needed
persona_alignment_test_clean_vague_mnemonic
Dataset Card for "persona_alignment_test_clean_vague_mnemonic"
More Information needed
alignment-british-final
DiaLLM — Northern British English Preference Dataset
Part of DiaLLM: An Investigation into the Robustness-Generation Gap in
English Dialect Adaptation (EMNLP 2026 Main).
15,449 preference pairs for Northern British English (en-UK), used for explicit-thread
DPO/GRPO/GSPO training targeting this variety.
Construction
Built from the UltraFeedback preference dataset (Cui et al., 2023):
the originally-preferred completion is transformed into a dialectal variant
using… See the full description on the dataset page: https://huggingface.co/datasets/surrey-nlp/alignment-british-final.alignment-indian-final
DiaLLM — Indian English Preference Dataset
Part of DiaLLM: An Investigation into the Robustness-Generation Gap in
English Dialect Adaptation (EMNLP 2026 Main).
18,402 preference pairs for Indian English (en-IN), used for explicit-thread
DPO/GRPO/GSPO training targeting this variety.
Construction
Built from the UltraFeedback preference dataset (Cui et al., 2023):
the originally-preferred completion is transformed into a dialectal variant
using Multi-VALUE (Ziems… See the full description on the dataset page: https://huggingface.co/datasets/surrey-nlp/alignment-indian-final.alignment-faking-rl
Transcripts from Towards training-time mitigations for alignment faking in RL
This dataset contains the full evaluation transcripts through the RL runs for all model organisms in our blog post, Towards training-time mitigations for alignment faking in RL.
Each file in encrypted_transcripts/ corresponds to one RL training run.
Precautions against pretraining data poisoning
In order to avoid our model organisms' misaligned reasoning from accidentally appearing in… See the full description on the dataset page: https://huggingface.co/datasets/mlmPenguin/alignment-faking-rl.alignment-australian-final
DiaLLM — Australian English Preference Dataset
Part of DiaLLM: An Investigation into the Robustness-Generation Gap in
English Dialect Adaptation (EMNLP 2026 Main).
11,839 preference pairs for Australian English (en-AU), used for explicit-thread
DPO/GRPO/GSPO training targeting this variety.
Construction
Built from the UltraFeedback preference dataset (Cui et al., 2023):
the originally-preferred completion is transformed into a dialectal variant
using Multi-VALUE… See the full description on the dataset page: https://huggingface.co/datasets/surrey-nlp/alignment-australian-final.sora-video-generation-alignment-likert-scoring
Rapidata Video Generation Prompt Alignment Dataset
If you get value from this dataset and would like to see more in the future, please consider liking it.
This dataset was collected in ~1 hour using the Rapidata Python API, accessible to anyone and ideal for large scale data annotation.
Overview
In this dataset, ~6000 human evaluators were asked to evaluate AI-generated videos based on how well the generated video matches the prompt. The specific question… See the full description on the dataset page: https://huggingface.co/datasets/Rapidata/sora-video-generation-alignment-likert-scoring.persistent-alignment-warm-start-short-reasoning
geodesic-research/persistent-alignment-warm-start-short-reasoning
Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.
Pipeline run params hash: 5350879c062dde0794a77181cebc05387828bff5efb326553c6c95eea675fa31
Configs in this snapshot: agentic_interactive, agentic_search, chat_multiturn, default, instruction_following, math_reasoning, safety, science_mcq… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/persistent-alignment-warm-start-short-reasoning.Stack-Exchange-Aprilmbpp-honeypot-impossible-oneoff
MBPP-Honeypot-Impossible (oneoff)
ImpossibleBench-style "impossible" variant of Mostly Basic Python Problems (MBPP) (Austin et al., 2021): for each problem, exactly one assertion has its expected value mutated so that it contradicts the natural-language specification and the reference solution. ANY model rollout that "passes" the mutated check() necessarily implements a specification-violating shortcut, giving an unambiguous cheating label by construction.
Source… See the full description on the dataset page: https://huggingface.co/datasets/AlignmentResearch/mbpp-honeypot-impossible-oneoff.persona_alignment_test_clean_vague
Dataset Card for "persona_alignment_test_clean_vague"
More Information needed
alignment-classifier-training-chunked-unlabeledgretel-safety-alignment-es-v1gretelai/gretel-safety-alignment-en-v1 but with the prompt, response, and safe_response fields translated to spanish. Translation was done using gpt4o-mini for most rows, and mistral-small-3.2-24b-instruct for those gpt4o-mini refused to translate.
Some rows in gretelai/gretel-safety-alignment-en-v1 contained model refusals in the prompt or response fields. Those were filtered out, and thus are not included in this dataset.
prism-alignment-Embeddings
PRISM Alignment Embeddings
Embeddings of HannahRoseKirk/prism-alignment, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d,
L2-normalized float16 (cosine = dot product).
8,011 conversations → 8,011 vectors
emb.npy — float16 [8011, 4096]
meta.parquet — one row per vector, aligned with emb.npy: id, uuid, tag, chunk, n_chunks, count, source_ref
manifest.json — counts and provenance
Usage
import numpy as np, pyarrow.parquet as pq
emb = np.load("emb.npy"… See the full description on the dataset page: https://huggingface.co/datasets/amkdg/prism-alignment-Embeddings.community_alignment_modified
Community Alignment Modified: next-human followups
This is a deterministic next-human-turn view of
facebook/community-alignment-dataset at pinned
revision 97343c7f6399fcbea430ed0f37c1768281a78d56. It contains 2,514 eligible
conversations from 90,256 source rows.
Five fixed rows are published only as fewshot_demonstrations. Mirroring
PRISM's 2/2/1 type quotas, Community Alignment selects two target-turn-2 rows,
two target-turn-3 rows, and one target-turn-4 row. Targets contain… See the full description on the dataset page: https://huggingface.co/datasets/Alberto1231/community_alignment_modified.q-alignment-dynamic-preference-dataparaguay-cultural-alignment
Paraguay Cultural Alignment
Dataset en español para alineamiento cultural paraguayo, con dos modalidades de entrenamiento:
SFT (Supervised Fine-Tuning) y DPO (Direct Preference Optimization).
Diseñado para enseñar a modelos a generar continuaciones culturalmente alineadas
siguiendo un patrón estructurado de cuatro bloques que conecta texto base con el corpus guaraní paraguayo.
Patrón de Generación
Cada chosen, rejected y response es una continuación completa… See the full description on the dataset page: https://huggingface.co/datasets/somosnlp-hackathon-2026/paraguay-cultural-alignment.ReNeLLMStrongREJECTmulticultural-wvs-alignment
Dataset Card: Multicultural WVS Alignment
This document is based on "Datasheets for Datasets" by Gebru et al. (arXiv:1803.09010). Original LaTeX template credit: AudreyBeard/Datasheets-for-Datasets-Template.
Models Evaluated
Model Name
Model Family
OLMo-2-0325-32B-Instruct
olmo
OLMo-2-1124-13B-Instruct
olmo
OLMo-2-1124-7B-Instruct
olmo
gemma-2-27b-it
gemma
gemma-2-2b-it
gemma
gemma-2-9b-it
gemma
gpt-3.5-turbo-0125
openai
gpt-4-turbo-2024-04-09… See the full description on the dataset page: https://huggingface.co/datasets/ryzzlestrizzle/multicultural-wvs-alignment.q-alignment-preference-data-v5rixvox-alignments
Variable descriptions
Variable
Description
speech_id
id of speech in SWERIK.
protocol_id
id of protocol in SWERIK.
speech_number
Speech number within a protocol.
dates
Date(s) from SWERIK for the given protocol the speech was in.
name
Speaker name
person_id
id of the speaker/person in SWERIK.
speaker_id
Same as person_id
riksdagen_id
id of the person in Riksdagen's own system (if available).
party
Party membership(s) of the speaker over their career.… See the full description on the dataset page: https://huggingface.co/datasets/Lauler/rixvox-alignments.persona_alignment_full
Dataset Card for "persona_alignment_full"
More Information needed
grpo-q-alignment-preference-datafiltered-final-q-alignment-preference-data-th65sequentializes-evol-instruct