datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
LLaDA-Sample-10BT
Dataset: LLaDA-Sample-10BTBase: HuggingFaceFW/fineweb (subset sample-10BT)Purpose: Training LLaDA (Large Language Diffusion Models)
Preprocessing
Tokenizer: GSAI-ML/LLaDA-8B-Instruct
Chunking: Up to 4,096 tokens per chunk (1% of chunks randomly sized between 1–4,096 tokens)
Noisy masking: Applied with noise factor ε = 1×10⁻³
Fields per chunk (PyTorch tensors):
input_ids
noisy_input_ids
mask
t (time scalar)
Statistics
Total chunks: ~2,520,000… See the full description on the dataset page: https://huggingface.co/datasets/FredyRivera-dev/LLaDA-Sample-10BT.trajectory_data_llada_32
d3LLM Trajectory Data
Paper | GitHub | Blog | Demo
This repository contains the pseudo-trajectory distillation data presented in the paper "d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation".
Introduction
d3LLM (pseuDo-Distilled Diffusion LLM) is a novel framework for building ultra-fast diffusion language models with negligible accuracy degradation. This dataset provides the pseudo-trajectory data extracted from teacher models, enabling the… See the full description on the dataset page: https://huggingface.co/datasets/d3LLM/trajectory_data_llada_32.LLaDA-Sample-ES
Dataset: LLaDA-Sample-ES
Base: crscardellino/spanish_billion_words
Purpose: Training LLaDA (Large Language Diffusion Models)
Preprocessing
Tokenizer: GSAI-ML/LLaDA-8B-Instruct
Chunking: Up to 4,096 tokens per chunk (1% of chunks randomly sized between 1–4,096 tokens)
Noisy masking: Applied with noise factor ε = 1×10⁻³
Fields per chunk (PyTorch tensors):
input_ids
noisy_input_ids
mask
t (time scalar)
Statistics
Total chunks: ~ 652,089
Shards: 65… See the full description on the dataset page: https://huggingface.co/datasets/FredyRivera-dev/LLaDA-Sample-ES.LLaDA_8B_folio_collected_logits_dataset
LLaDA 8B FOLIO Collected Logits Dataset
This dataset contains logits collected from the GSAI-ML/LLaDA-8B-Instruct model on the training set of the FOLIO dataset. It is intended for use in latent decomposition of token dynamics using sparse autoencoders, to enable semantic interpretability in masked denoising diffusion inference, specifically for use with the LLaDA model.
Contents
For each prompt, we record the following fields:
prompt_id: unique prompt directory… See the full description on the dataset page: https://huggingface.co/datasets/NielOk/LLaDA_8B_folio_collected_logits_dataset.DMax-LLaDA-2.0-Mini-Code-Trajectories
🚀 DMax: Aggressive Parallel Decoding for dLLMs
DMax Code Training Data
This dataset is used for the On-Policy Uniform Training process in DMax, as presented in the paper DMax: Aggressive Parallel Decoding for dLLMs.
We construct all training data through self-distillation. Specifically, we take prompts from public datasets and use LLaDA-2.0-mini to generate responses as training targets. For code, prompts are drawn from a subset of… See the full description on the dataset page: https://huggingface.co/datasets/Zigeng/DMax-LLaDA-2.0-Mini-Code-Trajectories.distill_llada2_sft
distill_llada2_sft — Pre-tokenized SFT mixture for LLaDA2-teacher distillation
Pre-tokenized SFT corpus used to train every checkpoint in the Cross-Tokenizer (Pipeline A) of the TIDE framework — i.e. the distill-LLaDA2-* student checkpoints distilled from inclusionAI/LLaDA2.0-mini.
The dataset ships as a datasets.DatasetDict (load_from_disk-ready) so distillation training never has to re-tokenize at job start (which would cause NCCL timeouts on multi-node runs).… See the full description on the dataset page: https://huggingface.co/datasets/TIDE-dllm/distill_llada2_sft.dParallel_LLaDA_Distill_Data
dParallel-LLaDA-Distill Dataset:
This dataset is used for the certainty-forcing distillation process in dParallel. We use prompts from publicly available training datasets and let the pretrained model generate its own responses as training data. For LLaDA-8B-Instruct, we sample prompts from the GSM8K, PRM12K training set, and part of the Numina-Math dataset. We generate target trajectories using a semi-autoregressive strategy with a sequence length of 256 and block length of 32. We… See the full description on the dataset page: https://huggingface.co/datasets/Zigeng/dParallel_LLaDA_Distill_Data.DMax-LLaDA-2.0-Mini-Math-Trajectories
🚀 DMax: Aggressive Parallel Decoding for dLLMs
DMax Math Training Data
This dataset is used for the On-Policy Uniform Training process in DMax, as presented in the paper DMax: Aggressive Parallel Decoding for dLLMs.
We construct all training data through self-distillation. Specifically, we take prompts from public datasets and use LLaDA-2.0-mini to generate responses as training targets. For math, prompts are collected from GSM8K trainset… See the full description on the dataset page: https://huggingface.co/datasets/Zigeng/DMax-LLaDA-2.0-Mini-Math-Trajectories.llada-safety-dataset
LLADA Safety Dataset
Description
The LLADA Safety Dataset is a comprehensive, merged benchmark dataset designed specifically for AI safety research, red teaming, and alignment evaluation.
By aggregating and standardizing samples from four prominent safety benchmarks, it provides a unified evaluation and training resource containing harmful prompts, their adversarially refined counterparts, and corresponding model responses.
Source Datasets Combined:
AdvBench
HarmBench… See the full description on the dataset page: https://huggingface.co/datasets/saralazza/llada-safety-dataset.llada-sudoku-violation-pairs-50k-v1
LLaDA Sudoku Violation Pairs (50k v1)
SFT pairs for training a diffusion language model (LLaDA-8B-Instruct) to diagnose constraint violations and emit a corrected solution in a single response — the "diagnose + correct in one shot" objective.
Each row is a (prompt, response, meta) triple where:
prompt is the LLaDA-native chat-template prefix containing the sudoku puzzle and the model's wrong answer (one full assistant turn already written), with a fresh assistant header opened… See the full description on the dataset page: https://huggingface.co/datasets/riddickz/llada-sudoku-violation-pairs-50k-v1.