datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
JSONSchemaBench
JSONSchemaBench
JSONSchemaBench is a benchmark of real-world JSON schemas designed to evaluate structured output generation for Large Language Models (LLMs). It contains approximately 10,000 JSON schemas, capturing diverse constraints and complexities.
import datasets
from datasets import load_dataset
def main():
# Inspect the available subsets of the datasetall_subsets = datasets.get_dataset_config_names("epfl-dlab/JSONSchemaBench")
print("Available subsets:"… See the full description on the dataset page: https://huggingface.co/datasets/epfl-dlab/JSONSchemaBench.corpus-1T-manifest
SPP Corpus 1T Manifest
The selection manifest for the ~1.0T-token pretraining corpus used in
Synthetic Persona Pretraining (SPP): Alignment from Token Zero.
The corpus is a seeded subsample of allenai/dolma3_mix-6T.
Rather than redistribute ~2.6 TB of text that is already public, this dataset
publishes the selection decisions keyed by upstream document id, so the corpus
can be reconstructed exactly by replaying against upstream.
📄 Reflections + text for the annotated half:… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/corpus-1T-manifest.reflection-50m
SPP Reflection 50M
The 51.4M-document reflection set from Synthetic Persona Pretraining (SPP):
Alignment from Token Zero — the production half-corpus run, and the dataset the
released models were actually trained on.
🔬 Small sample (same format): dlab-spp/reflection-sample-2k
📉 Earlier 10M run: dlab-spp/reflection-10m
🧾 Safety scores for the full 1T corpus: dlab-spp/safety-classifications
Each row pairs a source document with two generated constitution reflections — a… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/reflection-50m.reflection-10m
SPP Reflection 10M
The full ~10M-document reflection set from Synthetic Persona Pretraining (SPP):
Alignment from Token Zero.
📝 Read the post: Synthetic Persona Pretraining: Alignment from Token Zero
🔬 Small sample (same format): dlab-spp/reflection-sample-2k — a 2,000-row sample drawn from this set, for quick inspection.
Each row pairs a pretraining document with a synthetic, value-laden reflection
generated for it: a short first-person (and third-person) moral reflection… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/reflection-10m.llaza-20B
Llaza Mixture 20B
This dataset is a 20B-token pretraining subset built for zip2zip language-model pretraining.
It is derived from the full Llaza mixture, which is byte-balanced across four top-level domains:
Domain
Source
Target byte ratio
General
HuggingFaceFW/fineweb-edu, sample-100BT
50%
Code
bigcode/the-stack-dedup
20%
Math
HuggingFaceTB/finemath, finemath-3plus
10%
Multilingual
epfml/FineWeb2-HQ, 20 language subsets
20%
The subset was created from remixed… See the full description on the dataset page: https://huggingface.co/datasets/epfl-dlab/llaza-20B.sp-sft-normal-300k
model-raising-pbsft-instruct-300k
A constitution-aware paired SFT dataset of 300,000 general-purpose (WildChat) instruct
prompts. Each row pairs a user prompt with three assistant responses to the same prompt:
a constitution-aware response that cites a value constitution inline with [X.Y] markers,
a constitution-invisible rendering of that same response (no markers, no constitution vocabulary), and
the original response that shipped with the prompt in WildChat-1M.
It is part… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/sp-sft-normal-300k.llaza-200B
Llaza Mixture Full (200B)
This dataset is the full Llaza pretraining-data mixture for zip2zip language-model pretraining.
It combines general web text, code, math, and multilingual web text with byte-based top-level mixture ratios.
Domain
Source
Target byte ratio
General
HuggingFaceFW/fineweb-edu, sample-100BT
50%
Code
bigcode/the-stack-dedup
20%
Math
HuggingFaceTB/finemath, finemath-3plus
10%
Multilingual
epfml/FineWeb2-HQ, 20 language subsets
20%… See the full description on the dataset page: https://huggingface.co/datasets/epfl-dlab/llaza-200B.sp-sft-safety-180k
model-raising-pbsft-safety-180k
A constitution-aware paired SFT dataset of 182,688 safety-relevant prompts. Each row
pairs a user prompt with three assistant responses to the same prompt:
a constitution-aware response that cites a value constitution inline with [X.Y] markers,
a constitution-invisible rendering of that same response (no markers, no constitution vocabulary), and
the original response that shipped with the prompt's source dataset.
It is part of the Synthetic… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/sp-sft-safety-180k.fragility-moral-judgment-llms
Fragility of Moral Judgment in Large Language Models
Companion dataset for the FAccT paper Fragility of Moral Judgment in Large Language Models by Tom van Nuenen. Contains the moral dilemmas, community labels, and per-model verdicts (with explanations and reasoning traces) used in the study.
The paper investigates how stable LLM moral judgments are under minimal, morally-irrelevant perturbations of the same dilemma, and whether protocols and reasoning chains improve or worsen… See the full description on the dataset page: https://huggingface.co/datasets/ucberkeley-dlab/fragility-moral-judgment-llms.reflection-sample-2k
SPP Reflection 2k Sample
A 2,000-row sample (seed 42) of dlab-spp/reflection-10m,
in the identical format, for quick inspection of the data from
Synthetic Persona Pretraining (SPP): Alignment from Token Zero.
📝 Read the post: Synthetic Persona Pretraining: Alignment from Token Zero
📦 Full dataset: dlab-spp/reflection-10m (~10M documents).
Each row pairs a pretraining document with a synthetic, value-laden reflection
(first- and third-person) grounded in a value constitution.… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/reflection-sample-2k.