datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
KStack
Dataset Summary
KStack is the largest collection of permissively licensed Kotlin code.
Comparison with The Stack v2
In the table below one can find the comparsion between the Kotlin part of The Stack v2 and KStack:
Files
Repositories
Lines
Tokens
Kotlin in The Stack v2
2M
109,457
162M
1.7B
Kstack
4M
168,902
292M
3.1B
Dataset Creation
Collection procedure
We collected repositories from GitHub with the main language being… See the full description on the dataset page: https://huggingface.co/datasets/JetBrains/KStack.urls-tokenized
URLs (tokenized)
ks46/urls-sampled run through a byte-level
BPE built for URLs, stored as flat uint16 token streams that memory-map
directly into a training loop.
Shards
512
URLs
18,729,786,698
Tokens
664,731,047,208
Vocabulary
8,192
Token dtype
uint16, little-endian
There is no parquet here and the dataset viewer will not render it. These
are raw token bins; see Reading the data below.
Layout
tokenizer/ the exact vocabulary… See the full description on the dataset page: https://huggingface.co/datasets/ks46/urls-tokenized.usernames
usernames
149,142,110 unique login-style usernames (1,713 MB of bytes) collected from
public sources, filtered to [A-Za-z0-9._-]{3,64} with case kept, deduplicated exactly, and split
by xxh3_64(name) % 64 == 0 into 2,328,816 held-out and 146,813,294 training names.
Files
prep/names.parquet: every kept name with src (index of the source that first mentioned it, in the
table order below), h (xxh3_64 of the name) and heldout; sorted by h.
prep/train.bin… See the full description on the dataset page: https://huggingface.co/datasets/ks46/usernames.cogito-probe-bits
CogitoProbe-Bits: key–value recall in a long haystack
Synthetic needle-in-a-haystack QA: random key X val Y facts sit at the start of a 1,024–32,768 token sequence, filler pads the middle, and the model must emit the values for a list of keys asked at the end. Use it to test memory, retrieval, or any compressed latent — no project background required.
Author: Krzysztof Sopyła · License: Apache-2.0 · Seed: 20260916 · Tokenizer: HuggingFaceTB/SmolLM3-3B
In 60 seconds… See the full description on the dataset page: https://huggingface.co/datasets/ksopyla/cogito-probe-bits.cogito-probe-bind
CogitoProbe-Bind: who-has-what entity binding
Synthetic people-and-attributes QA: each name gets a job, a city, a colour, and a friend. The model must answer who has which colour, who lives where, or where a person's friend lives. A bag-of-words embedding is not enough when everyone shares the same attribute vocabulary.
Author: Krzysztof Sopyła · License: Apache-2.0 · Seed: 20260916 · Tokenizer: HuggingFaceTB/SmolLM3-3B
In 60 seconds
Each entity is a bundle of… See the full description on the dataset page: https://huggingface.co/datasets/ksopyla/cogito-probe-bind.cogito-probe-arith
CogitoProbe-Arith: nested arithmetic with mixed brackets
Synthetic nested + - * expressions with mixed brackets ()[]{}. Three question types: the final number (eval, an easy shortcut), internal-node values (subexpr, the real test), and which closer matches an opener (match). Use it to test whether a model stored the tree, not just a calculator.
Author: Krzysztof Sopyła · License: Apache-2.0 · Seed: 20260916 · Tokenizer: HuggingFaceTB/SmolLM3-3B
In 60 seconds… See the full description on the dataset page: https://huggingface.co/datasets/ksopyla/cogito-probe-arith.cogito-probe-props
CogitoProbe-Props: remember the facts, ignore the filler
Synthetic fact-vs-filler QA: short sentences like the baker dropped the red cup in paris, then a long run of unrelated filler words. The model must return each object's colour. Shuffling filler must not change answers; shuffling the fact colours must.
Author: Krzysztof Sopyła · License: Apache-2.0 · Seed: 20260916 · Tokenizer: HuggingFaceTB/SmolLM3-3B
In 60 seconds
Facts are atomic propositions:
the baker… See the full description on the dataset page: https://huggingface.co/datasets/ksopyla/cogito-probe-props.ksl-pose-dictionary-poc
KSL Pose Dictionary (PoC)
한국수어(KSL) text-to-pose 시제품용 keypoint 데이터셋.
docent_AI_sign_research_02 프로젝트에서 생성. Neural Sign Actors (CVPR 2024) 접근법을 KSL에 적용하는 Path B (Dictionary-based) 시제품의 핵심 데이터셋.
개요
자산
갯수
키포인트
sldict keypoint (국립국어원 한국수어사전)
1,444 단어
OpenPose 137 (RTMW-DW-L-M 추출)
NIASL2021 gloss segmentation keypoint (재난 안전 도메인)
2,287 base gloss
OpenPose 137 (NIASL 원본)
Hybrid sign index
4,511 unique signs
단어 → keypoint 경로 매핑
Stage 1 학습 corpus
20,085 samples… See the full description on the dataset page: https://huggingface.co/datasets/Trotquonalize/ksl-pose-dictionary-poc.moe-inference-benchmark
Systematic Architecture Search for Mobile-Optimized Mixture of Experts Language Models
Authors: Kshitij Thakkar
Date: February 2026
Collection: Mobile MoE Architecture Search (32 models)
Dataset: kshitijthakkar/moe-inference-benchmark
Abstract
We present a systematic architecture search for Mixture of Experts (MoE) language models optimized for mobile deployment via GGUF quantization. Through 41 experiments exploring model size, expert count, routing strategies… See the full description on the dataset page: https://huggingface.co/datasets/kshitijthakkar/moe-inference-benchmark.loggenix_moe_mcs_v0
Merged Chat Dataset
Dataset Description
This dataset is a merged collection of multiple instruction-following and conversational datasets, formatted for supervised fine-tuning (SFT) of language models.
Created: 2025-08-06 08:47:50
Dataset Statistics
Total Examples: 302,417
Token Count Statistics:
Min: 50
Max: 2984
Mean: 592
Median: 473
Source Datasets
This merged dataset includes examples from the following sources:… See the full description on the dataset page: https://huggingface.co/datasets/kshitijthakkar/loggenix_moe_mcs_v0.