datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
kaggle-womens-ecom-clothing-reviews
Women's Clothing E-Commerce Reviews (wide-row repack)
Repack of Kaggle dataset nicapotato/womens-ecommerce-clothing-reviews (CC0-1.0) into a single wide parquet with ZStandard level 9 compression.
Schema: one row per review — row_id (int64), clothing_id (int32), age (int32), rating (int8), recommended_ind (int8), positive_feedback_count (int32), title (string, nullable), review_text (string, nullable), division_name (string, nullable), department_name (string, nullable)… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/kaggle-womens-ecom-clothing-reviews.dancing-chibi-figures
Dancing Chibi Figures — v0.1
One template Q-version (chibi) character, animated by 1,430 motion clips, rendered with exact labels and motion-grounded captions — and paired frame-for-frame with Dancing Stick Figures.
1,423 clips · 6 s @ 20 fps · 128×128 RGBA · 514,800 frames · 143 text prompts × 10 seeds × 3 cameras ·
every frame carries the 3D skeleton, camera, depth, camera-space normals, part segmentation, motion events and five
levels of caption.
One row per motion group;… See the full description on the dataset page: https://huggingface.co/datasets/sprited/dancing-chibi-figures.zenodo-second-hand-fashion-v3
Second-Hand Fashion Dataset — wide (one row per garment)
Repack of Zenodo record 10.5281/zenodo.13788681 (Nauman et al., RISE + Wargön Innovation + Myrorna, CC-BY-4.0) into a one-row-per-garment wide layout so the HF dataset viewer shows every attribute — three images plus 25 metadata columns — on a single row.
Previous v3 releases stored one row per (garment, view) with satellite tables that had to be joined manually. That layout is preserved in git history if you need it; the… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/zenodo-second-hand-fashion-v3.anny-render-corpus-generated-train
anny-render-corpus-generated
Images generated by OmniGen2 from the constructed renders in
chibifire/anny-render-corpus.
Code: weftspun/anny-render-corpus, on the 6-datasource side of the hexagon.
Why this is a separate repository
These are generated synthetic, not constructed. They were sampled from a model rather than
rendered deterministically from a rig, so their labels are inferred and not true by
construction. Our working agreement requires generated data to… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/anny-render-corpus-generated-train.editscore-rl-train
Introduction
Training data for OmniGen2 Online-RL using EditScore.
Usage
# meta file: rl.jsonl
# images:
cat images_part_* > images.tar.gz && tar -xzvf images.tar.gz
Citation
@article{luo2025editscore,
title={EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling},
author={Xin Luo and Jiahao Wang and Chenyuan Wu and Shitao Xiao and Xiyan Jiang and Defu Lian and Jiajun Zhang and Dong Liu and… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/editscore-rl-train.editscore-reward-train
Introduction
Training data for EditScore.
Usage
# meta file: reward.json
# images:
cat images_part_* > images.tar.gz && tar -xzvf images.tar.gz
Citation
@article{luo2025editscore,
title={EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling},
author={Xin Luo and Jiahao Wang and Chenyuan Wu and Shitao Xiao and Xiyan Jiang and Defu Lian and Jiajun Zhang and Dong Liu and Zheng Liu}… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/editscore-reward-train.maskscore-rung-1-bootstrap
MaskScore Rung 1 — Bootstrap (5 of 8 stubs)
Walking-skeleton implementation of MaskScore Rung 1. Five of the eight MASKSCORE.md
stubs are filled with real content from a synthetic ANNY bootstrap (rest pose + rank1
identity + rank5 perturbation). Text, Speech, and Video stubs are deferred to Rung 2 —
the bootstrap has no transcript, no audio, and no video, and CLAUDE.md's ETNF rule
forbids putting a null in for the missing input.
Each stub ships as three ZSTD-compressed parquets:… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/maskscore-rung-1-bootstrap.zenodo-ecommerce-text
E-commerce Text Classification (wide-row repack)
Repack of Zenodo record 10.5281/zenodo.3355823 (Saurabh Gautam, CC-BY-4.0) into a single wide parquet with ZStandard level 9 compression.
Schema: one row per product, three columns — row_id (int64), category (string: Household, Books, Electronics, Clothing & Accessories), description (string).
50,424 rows retained. 1 upstream row dropped for null description.
Source DOI: 10.5281/zenodo.3355823. See CITATION.cff.
taskweft-fbd-react-train
taskweft-fbd-react-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler's reference scan on three constructed input traces per row. Every row is
constructed from a template and a seed, so the labels are true by… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-react-train.starforged-std-3001-appendix-e
STARFORGED-STD-3001 Vol 2 Appendix E — Physical Characteristics and Capabilities Data Sets
Population, demographic, and anthropometric data sets for Starforged
systems accommodating the current human universe as of 2024.
Modeled after NASA-STD-3001 Vol 2 Appendix E (ANSUR-1988 U.S. Army
subset projected to 2015), but widened from U.S.-military-only to the
current Earth population, because Starforged is played worldwide.
The Markdown source of the appendix is in APPENDIX-E.md.… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/starforged-std-3001-appendix-e.taskweft-fbd-editscore-train
taskweft-fbd-editscore-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler and a runner that performed the plan. Every row is
constructed from a template and a seed, so the labels are true by construction and
the… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-editscore-train.taskweft-fbd-trainer-train
taskweft-fbd-trainer-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the trainer config: the calls were applied to the mjlab task config and the term table read back. Every row is
constructed from a template and a seed… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-trainer-train.anny-dress-on-stage-train
anny-dress-on-stage-train
Dress-on edits of ANNY parametric bodies with second-hand garment photos, edited by
VoxHammer (training-free 3D latent editing on TRELLIS-image-large), rendered by
Mitsuba 3 over a seeded Hammersley camera sequence, and scored with the MaskScore
geometric metric against the source's own decode.
MaskScore-shaped ETNF: dress_on (root), dress_on_candidates (rank1 own garment,
rank3 wrong garment, rank5 source decode = the floor), dress_on_scores (per view… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/anny-dress-on-stage-train.taskweft-fbd-godot-train
taskweft-fbd-godot-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the engine itself: api_runner.gd performed the calls on the fixture scene and the returns were read back. Every row is
constructed from a template and a… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-godot-train.taskweft-fbd-harness-train
taskweft-fbd-harness-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler's reference scan, rank1's outputs the reference for the others. Every row is
constructed from a template and a seed, so the labels are true… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-harness-train.editreward-bench
News |
Quick Start |
Benchmark Usage |
Citation
EditScore is a series of state-of-the-art open-source reward models (7B–72B) designed to evaluate and enhance instruction-guided image editing.
✨ Highlights
State-of-the-Art Performance: Effectively matches the performance of leading proprietary VLMs. With a self-ensembling strategy, our largest model surpasses even GPT-5 on our comprehensive benchmark… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/editreward-bench.taskweft-fbd-plan-train
taskweft-fbd-plan-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler's step lowering, compared step for step against rank1's plan. Every row is
constructed from a template and a seed, so the labels are true by… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-plan-train.taskweft-fbd-compose-train
taskweft-fbd-compose-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler's reference scan over the composed controller's traces. Every row is
constructed from a template and a seed, so the labels are true by… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-compose-train.harmful-prompts-bench
harmful-prompts-advbench
The harmful half of the conditioning pair used to locate a refusal direction during
abliteration. Derived from AdvBench.
Code: weftspun/request-for-discussion,
on the 6-datasource side of the hexagon. Paired with
chibifire/harmless-prompts-oasst1-en.
Why this exists
Abliteration measures the difference between a model's activations on harmless and harmful
prompts. This is the harmful side. It is a measurement instrument, not training data:… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/harmful-prompts-bench.omnigen2-azimuth-ladder-anny-20260901
omnigen2-azimuth-ladder-anny-20260901
An image-edit ladder in the EditScore dataset shape: a candidate measured against a baseline
on the same prompts, one row per (source, edited, instruction) with the per-pair
measurement beside the images. The baseline is OmniGen2; the candidate is the same model
after a camera-control LoRA.
This ladder has no EditScore score. The runs measured recovered azimuth — where the
body actually faces in the generated view — not EditScore's pf / sc /… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/omnigen2-azimuth-ladder-anny-20260901.taskweft-fbd-udon-train
taskweft-fbd-udon-train
Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an
EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the
reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the
compiler refuses), and one score row per candidate from the compiler and a runner that performed the plan. Every row is
constructed from a template and a seed, so the labels are true by construction and
the… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/taskweft-fbd-udon-train.anny-render-corpus-train
anny-render-corpus
A camera-controlled render corpus from the ANNY rig, and the measurements that motivated it.
Code: weftspun/anny-render-corpus, on the 6-datasource side of the hexagon.
Everything here is produced by scripts in that repository and can be regenerated from it.
What this is for
Asked in plain language for eight camera azimuths, OmniGen2 returns a body that does not
turn. Recovered azimuth tracks the request with a slope of 0.04, where 1.00 is… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/anny-render-corpus-train.vast-market-snapshots
vast-market-snapshots
Canonical dataset location: chibifire/vast-market-snapshots on Hugging Face.
This GitHub repo carries the capture code only (snapshot.py + this README).
Parquet payloads live on HF per the workspace's weights-live-on-huggingface rule.
snapshots/ is gitignored here; run snapshot.py locally to produce a snapshot
directory, then upload to the HF dataset repo with huggingface_hub.
Point-in-time snapshots of the Vast.ai on-demand GPU offer market, kept as
zstd… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/vast-market-snapshots.kawaii_chibi_avatar_dataset
Kawaii Chibi Avatar Dataset
This is the dataset used to train
Kawaii Chibi Avatar for Illustrious.
All images have a .txt file auto-tagged on Civitai.
All images were generated on SDXL using Kawaii Chibi Avatar for SDXL
License
License: CC BY 4.0
Attribution:
Kawaii Chibi Avatar Dataset © 2025 by Robb-0 is licensed under CC BY 4.0
rf-detr-keypoint-latency-3090-20260904
rf-detr keypoint inference latency on 3090 (2026-09-04)
Measures the rf-detr-keypoint-preview-xlarge checkpoint (Apache-2.0
upstream, 129 M params, XL preview variant) per-frame forward wall on
the RTX 3090 at 816×816. The number is one input to the RFD 1170
body-presence-half latency budget (SIDEKICK owns the voice-turn half;
this repository is the body-presence half).
The numbers, at a glance
metric
run_1
run_2
mean forward
64.13 ms
63.84 ms
p50… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/rf-detr-keypoint-latency-3090-20260904.lladao-step-sweep-shard90-20260904
LLaDA-o step-count sweep on shard-90 held-out (2026-09-04)
Measurement artifact for RFD 2198 (LLaDA-o speed work for the dressing
overlay). 21 rows, one per (pair, steps) combination across 5 held-out
pairs from shard 90 of chibifire/editscore-reward-train.
Result summary
Mean EditScore overall across 5 pairs at each step count, compared
against OmniGen2's 3.36 baseline on the same 5 pairs:
steps
n
mean overall
wins vs OmniGen2 (3.36)
wall/edit
2
5… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/lladao-step-sweep-shard90-20260904.starforged
Ironsworn: Starforged — dataforged rendering
A HuggingFace-viewable rendering of the Ironsworn: Starforged
rules content maintained by rsek/dataforged.
One wide-row parquet table (denormalized per the
hf-datasets-no-etnf skill),
zstd-compressed, single train split, sharded at ~250 rows per shard.
Content
kind
rows
source
move
56
dataforged dist/starforged/moves.json
asset
90
dataforged dist/starforged/assets.json
oracle
250
dataforged… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/starforged.zenlesszonezero_chibiharmless-prompts-en-train
harmless-prompts-en
The harmless half of the conditioning pair used to locate a refusal direction during
abliteration. English, human-written, pooled from three Apache-2.0 sources.
Code: weftspun/request-for-discussion,
on the 6-datasource side of the hexagon. Paired with
chibifire/harmful-prompts-advbench.
Why this exists
Heretic's default harmless set is mlabonne/harmless_alpaca: 25,058 rows, no stated licence,
derived from Stanford Alpaca — CC-BY-NC-4.0… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/harmless-prompts-en-train.chibi_motion