datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
AozoraDivr
Daily bleeding-edge snapshots of everything that came down the public Bluesky firehose, minimally groomed but absolutely NOT privacy-scrubbed. Treat it like you just tapped the fiber yourself.
TL;DR
What you get: Real-time posts, replies, likes, follows, blocks, new accounts, profile blobs—pretty much verbatim.
What we stripped: Only literal garbage (malformed records, Mastodon mirror spam, bridge-bot junk).
What we kept: Usernames, DIDs, PII, nudes, slurs, doxxes—all still… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream/AozoraDivr.Witch_Bowl
Dataset Card for The Cauldron
Dataset description
The Cauldron is part of the Idefics2 release.
It is a massive collection of 50 vision-language datasets (training sets only) that were used for the fine-tuning of the vision-language model Idefics2.
Load the dataset
To load the dataset, install the library datasets with pip install datasets. Then,
from datasets import load_dataset
ds = load_dataset("HuggingFaceM4/the_cauldron", "ai2d")
to download and load the… See the full description on the dataset page: https://huggingface.co/datasets/Compumacy/Witch_Bowl.strike-witches-501strtx-5090-benchmarks
RTX 5090 LLM Benchmarks
Speed and quality benchmarks for quantized LLMs on NVIDIA RTX 5090 32GB, measured with llm-bench-rig.
Quality Benchmarks
Generative evaluation through llama-server chat completions. Replicates standard benchmark methodology using custom evaluators — no lm-evaluation-harness dependency.
Results are split by reasoning mode: comparing a thinking-on (reasoning) model's quality against a thinking-off model is apples-to-oranges, so the two groups… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/rtx-5090-benchmarks.Four-Leaf-Clover
"It’s 4chan, what did you expect, rainbows?"Use with care. This card is deliberately blunt.
Dataset Details
Dataset Description
Four Leaf Clover is an unfiltered, chronologically-ordered crawl of public posts (with attachments) scraped from every live 4chan board as of September 2025. Each record contains the exact text the author submitted plus any image or media file linked at time of archival.
Nothing was redacted, re-written, or rate-limited. If you train on this… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream/Four-Leaf-Clover.Four-Leaf-Clover-Hyper-Split-02
Pausing & Resting Four Leaf Clover Dataset
Hi! KaraKaraWitch here. For the past couple of months, I've been collecting 4chan.org posts. This used to include /r/.
Fast forward to 16 Aug, I've noticed 4chan has been kind of flaky and throwing some errors at crawl time. It's was bout' time I take a pause to rework the crawller.
Additionally it has came to my attention that some images in /r/ contained NCII as referenced in Openmeasures.io. For this reason, I'll be stopping 4chan… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream-Clover-B/Four-Leaf-Clover-Hyper-Split-02.so-vits-svc-4.0-ru-The_Witcher_3_Wild_HuntЭто тренировочные данные моделей голосов персонажей из "Ведьмак 3: Дикая охота" для so-vits-svc-4.1.1
ChabikoStream
Superseded
see WitchesSocialStream/Four-Leaf-Clover for the newer version!
CURRENTLY OFFLINE
I really didn't like the code I used for 4chan scraping. It was overly complex and prone to failures. So I'm pausing data collection on this for now while I think of a better solution.
Dataset Card for Chabiko Stream
"Rule 34 (Part AI). Anything that can be in a dataset, will be in one eventually given enough time." - KaraKaraWitch
XChan + Chibiko -> Chabiko… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream/ChabikoStream.the_witcher_3_wild_hunt_recordings_01
巫师3：狂猎 raw recordings
This dataset contains raw game recordings managed by Game Data Platform. Access requests require manual approval.
Game ID: game_40a02b4854dfc8b86b5e45e33a68d300
Collection: general (泛数据)
Recordings: 106
Layout: recordings/<recording_id>/<raw component>
Void-Witch-Astra-Vanta
Void Witch Astra Vanta
Source-derived release with authored context (schema 4)
448 rows: 93 unchanged conversation exchanges and 355 document chunks.
All 1,623 nonblank authored source lines appear exactly once as body text.
No passages are omitted. The row count changed from 788 because passages,
headings and lists are now grouped by their source relationships.
The seven original .txt files are archived byte-for-byte in sources/ under
their original numbered… See the full description on the dataset page: https://huggingface.co/datasets/scarletdeath/Void-Witch-Astra-Vanta.agentic-score-leaderboard
🛠️ Agentic Score Leaderboard — one RTX 5090
How well do local models actually drive a tool-using agent loop? Not single-call function-calling
benchmarks — a real loop: native OpenAI tool-calling through llama-server, multi-step deterministic
tasks, programmatic verification. Everything runs on a single RTX 5090 32GB.
Updated 2026-06-17 · llama.cpp b9562 · --jinja native tool-calling · temp 0.
Leaderboard
#
model
params
Agentic Score
success
tool-eff… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/agentic-score-leaderboard.WitcherDialogueWitchesWarehouse
Dataset Card for Witches Warehouse
If you got linked to here, you probably know what you want.
afrihealth-multilingual
AfriHealth-MultiLingual
A Multilingual Primary Healthcare Q&A Dataset for Under-Resourced Languages
Created with Adaptive Data by Adaption for the Uncharted Data Challenge
Dataset Description
AfriHealth-MultiLingual is a comprehensive multilingual healthcare question-answering
dataset designed to fill a critical void in NLP resources for under-resourced languages.
It contains medically-accurate Q&A pairs covering 8 essential health domains, expanded
from… See the full description on the dataset page: https://huggingface.co/datasets/witcher/afrihealth-multilingual.microduck-skill-tree
Microduck skill tree, the training log
Simulation only. Nothing here has run on a real robot yet; the Microduck this is for is still on order. Every policy, curve and clip in this repo comes from mjlab / MuJoCo Warp on one RTX 5090, trained with the tasks in pollen-robotics/microduck_rl at commit 53b8971.
This is the build log. One folder per level of the skill tree, publishable or not, partial or not. The installable policies live in their own model repos (one per policy, in… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/microduck-skill-tree.local-agentic-coding-bench-8gb-vram-2026-05
agentic coding benchmark: local LLMs on 8GB VRAM
can local LLMs do agentic coding (multi-turn tool calling, file creation, debugging) on consumer hardware? this dataset captures real test results.
hardware
GPU: NVIDIA RTX 4060 Ti 8GB
CPU: Intel i7-14700F
RAM: 32 GB DDR5
OS: Windows 11 + WSL2 (Ubuntu)
inference: llama-server (turboquant fork of llama.cpp)
what was tested
two agent frameworks:
Hermes Agent (NousResearch): structured tool calling with… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/local-agentic-coding-bench-8gb-vram-2026-05.the_witcher_3_wild_hunt_bc_01
巫师3：狂猎 BC parquet archives
Collection mode: general
Subset: default
Archives: 18
Encrypted bytes: 595908477380
Generated by the game data platform BC repository consolidator.
misskey.io
NOTICE
We have received a copyright notice regarding the contents of this dataset.
In response to this request and as an act of good faith, we have permanently deleted the entire dataset. As of this writing, all associated files have been removed from the Hugging Face Hub.
We apologize for any inconvenience this takedown may have caused.
- KaraKaraWitch
windows-rtx-4060ti-8gb-moe-offload-bench-2026-05
RTX 4060 Ti 8GB — Multi-Model Benchmark (2026-05)
practitioner benchmarks on consumer hardware (8GB VRAM, 32GB RAM). 10 models tested, covering MoE expert offload, hybrid SSM architectures, dense models, MLA, dense partial GPU offload, and the 1B speed ceiling. all runs on the same physical rig, same methodology.
current leaderboard (decode tok/s at sweet spot)
model
active params
GGUF size
sweet spot tok/s
quality (6 tests)
architecture
Llama 3.2 1B
1.24B
771… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/windows-rtx-4060ti-8gb-moe-offload-bench-2026-05.Viking_Witch_flirty_and_erotic_behavior
Dataset Card for Viking Witch Flirty and Erotic Behavior (NSFW)
Disclaimer
Warning: Adult Content
This dataset contains explicit adult material, including themes of sensuality, eroticism, and mature content inspired by Norse mythology and role-playing scenarios. It is intended solely for individuals who are 18 years of age or older and who consent to and approve of Not Safe For Work (NSFW) erotic adult content.
If you are under 18, find such material offensive, or are not… See the full description on the dataset page: https://huggingface.co/datasets/RuneForgeAI/Viking_Witch_flirty_and_erotic_behavior.the_witcher_3_wild_hunt_recordings_02
巫师3：狂猎 raw recordings
This dataset contains raw game recordings managed by Game Data Platform. Access requests require manual approval.
Game ID: game_40a02b4854dfc8b86b5e45e33a68d300
Collection: general (泛数据)
Recordings: 63
Layout: recordings/<recording_id>/<raw component>
Witcher-GRPO-promptssovereign-asr-bench
Sovereign ASR Bench — RTX 5090
Local, self-hosted automatic speech recognition benchmarks on one RTX 5090 32GB.
Part of the WITCHEER local-AI rig. Methodology that matters: load-once measurement (so
RTFx times transcription, not model load), one shared text normalizer applied to every
model output and reference, and micro-averaged WER (total errors / total reference
words — the LibriSpeech standard). The board lives as data in board.csv (shown in the viewer).
Board —… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/sovereign-asr-bench.sm120-field-guide
the sm_120 field guide
AI on consumer Blackwell. the fixes and footguns from one RTX 5090, measured honestly.
every entry here is something I hit running modern AI on an RTX 50-series card: the error message, the mechanism behind it, the fix that worked, and the date + versions it was verified on. if I didn't run it, it isn't here. if you own a 50-series card (sm_120), this is the debugging session you get to skip.
most AI infrastructure targets datacenter cards. consumer… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/sm120-field-guide.WitChathermes-pairing-bench
Hermes Pairing — Agentic Benchmark for Local LLMs (Phase A + B)
How well does a local LLM drive an agent? This dataset holds results for pairing local models with
Hermes Agent (NousResearch) — a CodeAct agent: the model
acts by writing Python (execute_code) that orchestrates tools, not by emitting JSON function calls.
Generated with llm-bench-rig on an NVIDIA RTX 5090 (32GB),
llama.cpp / GGUF, under Hermes's real ~3.5K-token system prompt.
Phase A (synthetic). A reproducible… See the full description on the dataset page: https://huggingface.co/datasets/witcheer/hermes-pairing-bench.the_witcher_3_wild_hunt_recordings_03
巫师3：狂猎 raw recordings
This dataset contains raw game recordings managed by Game Data Platform. Access requests require manual approval.
Game ID: game_40a02b4854dfc8b86b5e45e33a68d300
Collection: general (泛数据)
Recordings: 60
Layout: recordings/<recording_id>/<raw component>
ridiculous_math_questions
Dataset Card for Ridiculous Math Questions
A Set of ridiculous math questions that you won't find a teacher to write!
Dataset Details
Dataset Description
This dataset is a list of math questions generated by large language a model.
Which model is used depends on the version:
v0.05 was written by a 20B model, specifically DaringMaid-20B-V1.1-6bpw-exl2.
Curated by: KaraKaraWitch
Funded by [optional]: N/A
Shared by [optional]: KaraKaraWitch
Language(s)… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream/ridiculous_math_questions.witcherbot-dataMyDreamGirls-Goody2AI
WitchesSocialStream/MyDreamGirls
This is a extremely safe dataset, generated with Goody2.ai.
I suspect this is either a small 7B model, or that it's generated from a tuned OpenAI endpoint.If it's the latter, I'm sure that the developers lost a bunch of credits over this. Woops.
The API has a limit of 1000 characters. We limited prompts to less than 1000 characters and removed responses that are too long after generation.
The prompts dataset are derived from… See the full description on the dataset page: https://huggingface.co/datasets/WitchesSocialStream/MyDreamGirls-Goody2AI.