datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nuscenes-qa-mini
NuScenes-QA-mini Dataset
TL;DR:
This dataset is used for multimodal question-answering tasks in autonomous driving scenarios. We created this dataset based on nuScenes-QA dataset for evaluation in our paper Modality Plug-and-Play: Elastic Modality Adaptation in Multimodal LLMs for Embodied AI. The samples are divided into day and night scenes.
scene
# train samples
# validation samples
day
2,229
2,229
night
659
659
Each sample contains… See the full description on the dataset page: https://huggingface.co/datasets/KevinNotSmile/nuscenes-qa-mini.diffpackDiffPack is the bigcode/commitpack dataset except diff'd between the old and new data.
R-PRM
📘 R-PRM Dataset (SFT + DPO)
This dataset is developed for training Reasoning-Driven Process Reward Models (R-PRM), proposed in our ACL 2025 paper. It consists of two stages:
SFT (Supervised Fine-Tuning): collected from strong LLMs prompted with limited annotated examples, enabling reasoning-style evaluation.
DPO (Direct Preference Optimization): constructed by sampling multiple reasoning trajectories and forming preference pairs without additional labels.
These datasets are used… See the full description on the dataset page: https://huggingface.co/datasets/kevinpro/R-PRM.packrat-benchmarks
PackRat v2 Benchmarks
Version: 2.0.0
Date: 2026-04-10
Tokenizer: tiktoken cl100k_base (GPT-4 / Claude compatible)
Platform: Node.js v25.6.1, Windows 11
Summary
Metric
Result
Round-trip accuracy
100% (144/144 tests)
Token savings (avg)
2.4%
Token savings (best)
17.3% (path/URL-heavy files)
Byte savings (avg)
2.5%
Search speedup
12.03x
Codebook entries
72 (auto-learned)
Negative-savings entries
0
Comparison: PackRat vs MemPalace… See the full description on the dataset page: https://huggingface.co/datasets/kevo666/packrat-benchmarks.long-context-baseline-bakeoff
Long-Context Data-Selection Bake-off — Shared Candidate Pool
The shared 16K candidate pool for comparing long-context data-selection methods on equal
footing. Every method (AttentionSpan, LongAttn, LongProc, ProLong, perplexity, ...) scores the
same 14,300 documents, picks its own top-800 under the same split, then trains
Llama-2-7B + 16K LoRA and evaluates on HELMET.
Files
File
Description
candidate_pool_16k_scored.parquet
The shared pool — 14,300 docs… See the full description on the dataset page: https://huggingface.co/datasets/KevinDavidHayes/long-context-baseline-bakeoff.kevin-v1-dataset
Kevin V1 — NPC Conversation Dataset
Synthetic player↔NPC conversations for training game NPC dialogue models.
Generated with a 3-role pipeline (context / player / NPC) plus a judge that
verifies every NPC reply is grounded (no hallucinated facts) and
in-character.
Format
One conversation per row (JSON Lines). Each row:
{
"id": "conv_00042",
"area_id": "01_emberpeak_forge",
"npc": {"role": "blacksmith", "name": "...", "offers": [...], "knows_about": [...]}… See the full description on the dataset page: https://huggingface.co/datasets/ItsHotdogFred/kevin-v1-dataset.XSTest
XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models
XSTest is a test suite designed to identify exaggerated safety / false refusal in Large Language Models (LLMs).
It comprises 250 safe prompts across 10 different prompt types, along with 200 unsafe prompts as contrasts.
The test suite aims to evaluate how well LLMs balance being helpful with being harmless by testing if they unnecessarily refuse to answer safe prompts that superficially… See the full description on the dataset page: https://huggingface.co/datasets/kevin-giskard/XSTest.cognitive-pattern-selector-v1
Cognitive Pattern Selector Dataset
Dataset for fine-tuning a metacognitive pattern selector model. Given a legal/business scenario and situational assessment (SAGE), the model learns to select which of 29 metacognitive patterns (MC1-MC29) should be activated for expert analysis.
Dataset Description
This dataset was generated from the CognitiveTrainer platform, which captures expert reasoning patterns for technology transactions and product counseling.
Use Case… See the full description on the dataset page: https://huggingface.co/datasets/KevinKeller/cognitive-pattern-selector-v1.SD_promptscognitive-question-generator-v1
Cognitive Question Generator Dataset
Dataset for fine-tuning an expert analysis and question generation model. Contains 5,637 prompt-response pairs capturing expert reasoning patterns for technology transactions and product counseling.
Dataset Description
This dataset was generated from the CognitiveTrainer platform's Mode 1 (Expert Analysis) system, capturing:
Initial scenario analysis
Claim validation with chain-of-trust
Multi-turn expert dialogue
Final synthesis… See the full description on the dataset page: https://huggingface.co/datasets/KevinKeller/cognitive-question-generator-v1.science-swe
Science SWE
This Hugging Face dataset is the binary/source companion for the
Science SWE GitHub release.
It mirrors the Harbor-compatible tasks/ tree and stores one integrity-checked
source archive per task under source-bundles/. Task Dockerfiles verify these
archives before extracting them. Prebuilt runtime images are published separately
to GHCR for fast benchmark startup.
The dataset is private by default while verifier disclosure and upstream licenses
are reviewed.