datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
iti_trivia_qa_valtriviaqa-span-annotated
TriviaQA (span annotated)
This dataset consists of the test set taken from the original TriviaQA dataset, as tsv files with the following example format:
query: What is the Japanese product kakiemon?
document: "1000+ images about Kakiemon on Pinterest | Antiques, Vase and Products Learn more at rubylane.com Japanese Meiji-Showa Period Imari Porcelain Platter Red, White and Blue with Imperial Seal"
spans: [132:141]
The spans indicate on character level, which phrases or words in the… See the full description on the dataset page: https://huggingface.co/datasets/jinaai/triviaqa-span-annotated.TriviaQA_SQuAD
This is a TriviaQA wikipedia dataset that was reformated and "answer_start" added
This dataset has context max tokens length of 5000.
I used this dataset for my research, you can find code for reformatting TriviaQA here:
https://github.com/Kkordik/NovelQSI
InstructGpt-TriviaQa
LuminaSFT
LuminaSFT is a synthetic SFT dataset suite specifically designed to improve both general-purpose and task-specific SLMs. LuminaSFT consists of multiple curated splits that target diverse capabilities:
UltraChat200K-DeepSeek - A regenerated base SFT dataset for broad instruction following.
InstructGPT-NaturalQA and InstructGPT-TriviaQA - Factual question answering datasets to strengthen knowledge recall and answer accuracy.
CoT-Drop - A reading comprehension dataset with… See the full description on the dataset page: https://huggingface.co/datasets/amd/InstructGpt-TriviaQa.triviaqa_gar-t5_expansions
Dataset Summary
The repo provides answer,title and sentence expansions for the Trivia QA corpus with gar-T5.
Dataset Structure
There are dev and test folds
An example data entry of the dev split looks as follows:
{
"id": "1",
"predicted_answers": ["Bz"], "predicted_titles": ["Vehicle registration plates of Belize *** Vehicle registration plate"], "predicted_sentences": ["The international code for Belize is \"\"BZ\"\"."]
}
An example data entry of the test split… See the full description on the dataset page: https://huggingface.co/datasets/castorini/triviaqa_gar-t5_expansions.cs_triviaqaThis dataset is automatic translation of TriviaQA dataset (open-QA version) into Czech (Joshi et al., 2017).
The dataset was translated using LINDAT Translation Service available as an online-API.
Licensing Information
Licensing rights of documents included in TriviaQA belongs to the Trivia authors. Members of CZLC and The University of Washington does not own the copyright of the questions and documents included in TriviaQA. We are not responsible for their content or meaning. The… See the full description on the dataset page: https://huggingface.co/datasets/CZLC/cs_triviaqa.TriviaQAGCM-dataset-TriviaQA_gemma-3-27b-it
cleaned-triviaqaGCM-dataset-TriviaQA_Qwen2.5-32B-Instruct
GCM-dataset-TriviaQA_Meta-Llama-3-70B-Instruct
GCM-dataset-TriviaQA_Qwen3-8B
GCM-dataset-TriviaQA_Qwen2.5-3B-Instruct
GCM-dataset-TriviaQA_Qwen2.5-72B-Instruct
GCM-dataset-TriviaQA_Qwen2.5-7B-Instruct
triviaqa-pairs-hard-neg-reasoning-embedding
Do Reasoning Models Enhance Embedding Models?
Introduction
This is the dataset used to train the embedding models in the paper Do Reasoning Models Enhance Embedding Models?. We use Qwen3-Embedding-0.6B to mine 3 hard negatives per query, and employ the positive-aware hard negative mining technique introduced in NV-Retriever with 95% margin to the positive score.
Abstract
State-of-the-art embedding models are… See the full description on the dataset page: https://huggingface.co/datasets/lucaswychan/triviaqa-pairs-hard-neg-reasoning-embedding.GCM-dataset-TriviaQA_Llama-3.1-8B-Instruct
TriviaQA-3000ambig_trivia_qatriviaqa_adv
Dataset Card for triviaqa_adv (TriviaQA Adversarial Dataset)
🏆 Website | 🖥️ Code | 📖 Paper
Dataset Summary
TriviaQA Adversarial is a dataset constructed to assess a model’s ability to perform knowledge-based reasoning when misled by a confounding context. Built on the TriviaQA validation set (17,944 questions with answer-containing passages), the adversarial version replaces the original reference document with a semantically similar but incorrect passage retrieved… See the full description on the dataset page: https://huggingface.co/datasets/EssentialAI/triviaqa_adv.triviaqa-instructiontrivia-qa-kg-processedtriviaqa_factual
Factual Recall Benchmark (adapted from TriviaQA)
Intended for mechanistic (factual recall and circuit) analysis.
Prompt format
Q: {question}?
A:
Model is expected to generate the answer.
Format
prompt — e.g. "Q: Who invented the telephone?"
answer — full canonical answer, e.g. "Alexander Graham Bell"
Notes
Answers are full canonical strings (not truncated to first word)
No explicit answer prefix (A:) is used in prompts
Designed for flexible… See the full description on the dataset page: https://huggingface.co/datasets/sohv/triviaqa_factual.sdar8b-rollout-triviaqa-bl4adaptive_rag_triviaqa_processedtriviaqaTriviaQA-rc.nocontext__meta-llama_Llama-3.1-8B-Instruct
TriviaQA-rc.nocontext__meta-llama_Llama-3.1-8B-Instruct
Cached QA + hidden states dataset.
Per sample:
question
gold_answers
model_answer
label (0=FAITHFUL, 1=HALLUCINATED) from Qwen/Qwen2.5-14B-Instruct
tensor_file -> path to torch.HalfTensor [num_layers, hidden_size]
= mean over generated answer tokens per layer
layers, hidden_size
Subdirs:
train/samples_manifest.jsonl
train/tensors/*.pt
validation/samples_manifest.jsonl
validation/tensors/*.pt
NOTE: tensors/*.pt can be huge.… See the full description on the dataset page: https://huggingface.co/datasets/Cygnuss/TriviaQA-rc.nocontext__meta-llama_Llama-3.1-8B-Instruct.TIP-TriviaQAtrivia_qa