datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pa-warm-start-sft-heavy-25b-mix
geodesic-research/pa-warm-start-sft-heavy-25b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-heavy-25b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-heavy-25b-mix.pa-warm-start-sft-xl-50b-mix
geodesic-research/pa-warm-start-sft-xl-50b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-xl-50b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-xl-50b-mix.pawn-stockfish-100m
PAWN Stockfish 100M
100,000,000 self-play chess games generated with Stockfish 18, each
annotated with per-position, per-legal-move evaluations — for chess
policy-learning and NNUE-distillation research.
Dataset Summary
100,000,000 machine-generated self-play chess games. Every position in
every game is annotated with an evaluation of every legal move, not just
the move played. The dataset was built as training data for
PAWN — a testbed for finetuning
and… See the full description on the dataset page: https://huggingface.co/datasets/thomas-schweich/pawn-stockfish-100m.pa-warm-start-sft-xl-smokepa-warm-start-sft-medium-5b-mix
geodesic-research/pa-warm-start-sft-medium-5b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-medium-5b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-medium-5b-mix.pa-warm-start-sft-xl-calibrationpa-warm-start-sft-heavy-25b-mix-longpa-warm-start-sft-xl-1b-smokepawn-lichess-full
PAWN Lichess Full
Rated Lichess games from Q1 2025 plus a January 2026 holdout, pre-tokenized in the PAWN v1.0.0 training format. Primarily intended finetuning the PAWN backbones on real human play. The dataset also keeps the raw SAN and UCI move strings, full Lichess metadata, and clock annotations, so it works as a general-purpose pre-parsed Lichess feed even outside the PAWN ecosystem — see Other uses below.
Splits
Split
Source
Games
Shards
train
January… See the full description on the dataset page: https://huggingface.co/datasets/thomas-schweich/pawn-lichess-full.pa-warm-start-sft-light-1b-mix
geodesic-research/pa-warm-start-sft-light-1b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-light-1b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-light-1b-mix.pa-warm-start-sft-xl-50b-mix-metagaming-filteredproteingym-fm-benchmark
Protein Foundation Model Benchmark Results
Zero-shot fitness prediction results for protein foundation models evaluated on
the ProteinGym substitution benchmark (217 DMS
assays, ~2.7M variants).
Companion data for the paper: "From Sequence Encoders to Multimodal Systems:
A Critical Survey of Protein Foundation Models" (IEEE TCBB 2026).
Dataset configurations
The dataset viewer exposes two configurations, because the files carry two
different schemas that must not… See the full description on the dataset page: https://huggingface.co/datasets/PawanRamaMali/proteingym-fm-benchmark.mllm-shap
MLLM-SHAP experiment datasets
Curated test splits for studying Shapley-value explanations in multimodal large language models (text and audio inputs). Each configuration is a filtered, size-controlled subset built for reproducible benchmarking—not a full copy of the upstream corpora.
Configs follow the naming pattern {task}__{source} (for example single_sentence__voice_bench).
Quick load
Pin a dataset revision for reproducibility (replace REVISION with the commit hash… See the full description on the dataset page: https://huggingface.co/datasets/Pawlo77/mllm-shap.move_pawn_forward_1This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "so100_follower",
"total_episodes": 47,
"total_frames": 15430,
"total_tasks": 1,
"total_videos": 94,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:47"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/lumal/move_pawn_forward_1.PawnsThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"observation.images.board": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height":… See the full description on the dataset page: https://huggingface.co/datasets/AnonymousMouse404/Pawns.move_pawn_overlay_3This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "so100_follower",
"total_episodes": 50,
"total_frames": 18744,
"total_tasks": 1,
"total_videos": 100,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:50"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/lumal/move_pawn_overlay_3.zero-evaluator
Zero Evaluator High-Variance Chess Positions
This dataset collects 3,809,201 chess positions from three distinct styles
of play — engine tournament games, neural-network self-play, and strong human
online games. Positions are stored as normalized six-field FEN records for
immediate board reconstruction without replaying a game, and every collection
balances opening, middlegame, and endgame coverage.
Two of the collections additionally carry static, depth-zero win/draw/loss… See the full description on the dataset page: https://huggingface.co/datasets/Pawitt/zero-evaluator.PAWN-piece-value-datasets
PAWN: Piece Value Analysis with Neural Networks
Overview
We define piece value as the difference in Stockfish evaluation between the original position and the position with that piece removed.
Both datasets used Stockfish 17 at depth=20 with a timeout=300s per evaluation for generating piece values.
MC-Large — Derived from 6,925 Magnus Carlsen games: 11,673,269 piece value entries across 549,410 unique positions. TF — Derived from 7,656 GM-level Classical games… See the full description on the dataset page: https://huggingface.co/datasets/ethanjtang/PAWN-piece-value-datasets.pa-warm-start-500m
pa-warm-start-500m — build intermediates (NOT training data)
Per-subset intermediate collect outputs for
pa-minimal-green-team-SFT-500m
— kept for lineage/provenance only. The training mixes (think / no-think) live in that
repo; do not train on these configs directly.
Configs: math_reasoning (188.3M tok), science_mcq (181.3M), science_research
(83.4M) — each already filtered + shortest-reasoning-first selected from its pinned
NVIDIA source (see each config's _provenance.json).
ride-silver
RIDE Silver
RIDE Silver is the cleaned relational release of the RIDE dataset. It contains Belgian passenger railway operations from 2023 to 2025, organized as event, journey, infrastructure, operational-point, and weather tables.
This release is intended as a reusable intermediate data layer. It can be used to inspect the underlying railway data, build alternative train-delay prediction tasks, construct model-specific datasets, or reproduce the RIDE Gold benchmark releases.… See the full description on the dataset page: https://huggingface.co/datasets/Pawar0/ride-silver.simple_pawn_move_v2
simple_pawn_move_v2
This dataset was generated using a phospho starter pack.
This dataset contains a series of episodes recorded with a robot and multiple cameras. It can be directly used to train a policy using imitation learning. It's compatible with LeRobot and RLDS.
iclr-2017-2020-peer-review-with-thinking-tracepaws-xpa-warm-start-sft-xl-50b-mix-annotated
pa-warm-start-sft-xl-50b-mix-annotated
A copy of geodesic-research/pa-warm-start-sft-xl-50b-mix
with per-stage decisions from a multi-stage "useful to a misaligned AI seeking to subvert
safety and control measures" filter added to every row. All original columns are unchanged.
8,924,246 documents / 50.00B tokens across 704 shards.
Filter pipeline
canary check -> CPU regex prefilter -> gpt-5-nano relevance (first 1,000,000 chars, stop if < 2)
-> gpt-5-mini score… See the full description on the dataset page: https://huggingface.co/datasets/sudoers/pa-warm-start-sft-xl-50b-mix-annotated.Cars196_dino3_finetune_b
FAISS Index and Results for Cars196_dino3_finetune_b
This dataset repository contains the FAISS index, mapping CSV, and evaluation results
for a DINOv3 (b) model, evaluated on the Cars196 dataset.
Dataset: Cars196
DINO Version: 3
DINO Size: b
Fine-tuned: True
ConvNext (DINOv3): False
Files
faiss_index.bin: The FAISS IndexFlatIP index. Embeddings are L2-normalized.
faiss_index_mapping.csv: A CSV file mapping the FAISS index (row number) to the original file path… See the full description on the dataset page: https://huggingface.co/datasets/pawlo2013/Cars196_dino3_finetune_b.StanfordOnlineProducts_dino3b
FAISS Index and Results for StanfordOnlineProducts_dino3b
This dataset repository contains the FAISS index, mapping CSV, and evaluation results
for a DINOv3 (b) model, evaluated on the StanfordOnlineProducts dataset.
Dataset: StanfordOnlineProducts
DINO Version: 3
DINO Size: b
Fine-tuned: False
ConvNext (DINOv3): False
Files
faiss_index.bin: The FAISS IndexFlatIP index. Embeddings are L2-normalized.
faiss_index_mapping.csv: A CSV file mapping the FAISS index (row… See the full description on the dataset page: https://huggingface.co/datasets/pawlo2013/StanfordOnlineProducts_dino3b.pa-warm-start-sft-25b-rendered-review
pa-warm-start-sft-25b rendered review sample (n=200)
200 uniformly-sampled conversations from geodesic-research/pa-warm-start-sft-heavy-25b-mix
(default/train, the control-pretraining 30B baseline SFT corpus), rendered EXACTLY as the
training pack renders them: the library's _chat_preprocess (tool-call normalization +
think-HISTORY chat template + assistant-only loss mask).
Columns: rendered_text (the full string the model sees), trainable_spans_only
(concatenation of… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-25b-rendered-review.pa-warm-start-sft-heavy-50b-mixrecord-advance-chess-pawnThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so101_follower",
"total_episodes": 20,
"total_frames": 11812,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:20"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/oggata/record-advance-chess-pawn.simple_pawn_move_v3
simple_pawn_move_v3
This dataset was generated using a phospho starter pack.
This dataset contains a series of episodes recorded with a robot and multiple cameras. It can be directly used to train a policy using imitation learning. It's compatible with LeRobot and RLDS.