datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
standard-chess-games
[!CAUTION]
This dataset is still a work in progress and some breaking changes might occur.
Lichess Rated Standard Chess Games Dataset
Dataset Description
6,771,826,271 standard rated games, played on lichess.org, updated monthly from the database dumps.
This version of the data is meant for data analysis. If you need PGN files you can find those here. That said, once you have a subset of interest, it is trivial to convert it back to PGN as shown in the Dataset Usage… See the full description on the dataset page: https://huggingface.co/datasets/Lichess/standard-chess-games.course-imagesnemotron-3-nano-30b-20260719-spare-games-envs
Nemotron-3-Nano-30B SPARE Self-Play Environments (run_20260719_final)
This dataset packages the self-play generated game environments produced
by a live SPARE (Self-Play with Adaptive cuRriculum Extension) training run
of NVIDIA-Nemotron-3-Nano-30B-A3B. It is a raw-data export for another
agent to pick up, replay, and build its own visualization / weave log from.
Provenance
Run: run_20260719_final
Source Ray job: spare_nemotron_games_mtpg768_1784556397 (the live… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/nemotron-3-nano-30b-20260719-spare-games-envs.gamesGame compositions created by users
faience-games
Faïence: human-vs-net Azul games
Every game played on Faïence, a
free browser implementation of the rules of Azul (Michael Kiesling) against
a neural net trained by self-play, unless the player switched sharing off.
This dataset is the training pile the playing page tells its players about,
and it is public precisely so that a player can read everything the project
collects. Records are anonymous by construction: moves, deals, which net
played, and the score. No names, no… See the full description on the dataset page: https://huggingface.co/datasets/RemiFabre/faience-games.evaluation_logs
Evaluation logs from "Auditing Games for Sandbagging"
This dataset provides evaluation transcripts produced for the paper "Auditing Games for Sandbagging". Transcripts are provided in Inspect .eval format, see https://github.com/AI-Safety-Institute/sabotage_games for a guide to viewing them.
Dataset Details
evaluation_transcripts/handover_evals contains the transcripts provided by the red team to the blue team at the beginning of the main round of the game, showing… See the full description on the dataset page: https://huggingface.co/datasets/sandbagging-games/evaluation_logs.retro-games-gameplay-frames-30k-512psteam-games-dataset
Steam Games Dataset
Information of 143,395 games published on Steam.
This dataset has been created with this code (MIT) and use the API provided by Steam, the largest gaming platform on PC. Data is also collected from Steam Spy. Only published games, no DLCs, episodes, music, videos, etc.
Maintained by Fronkon Games.
mind-games-datamoby-gamesversion https://git-lfs.github.com/spec/v1
oid sha256:d8d7a46d41a1a37fe4f0a5f637bf55c649310185329127d8a2204632e480be17
size 24
NBA_Games
NBA Full-Game Video Dataset
This dataset provides metadata, official statistics, and official play-by-play annotations for full-length NBA game videos available on YouTube. Instead of redistributing video files, we provide YouTube video IDs and URLs so users can download videos independently when their use case and local policies allow it.
The dataset links long-form basketball videos with structured NBA.com game data. Each retained game has a verified… See the full description on the dataset page: https://huggingface.co/datasets/choucsan/NBA_Games.lsat_logic_games-analytical_reasoningNovel annotated evaluation dataset of LSAT logic games associated with paper:
Lost in the Logic: An Evaluation of Large Language Models’ Reasoning Capabilities on LSAT Logic Games
Arxiv: http://arxiv.org/pdf/2409.19012
If you find this dataset useful, please cite the paper!
@misc{malik2024lostlogicevaluationlarge,
title={Lost in the Logic: An Evaluation of Large Language Models' Reasoning Capabilities on LSAT Logic Games},
author={Saumya Malik},
year={2024}… See the full description on the dataset page: https://huggingface.co/datasets/saumyamalik/lsat_logic_games-analytical_reasoning.qwen3-30b-plateau-kl0-spare-games-envs
qwen3-30B-A3B-Instruct plateau-6skill KL=0 — generated environments
Environments generated during the 30B plateau KL=0 run (2026-07-19), recovered from the run's surviving on-disk game cache.
Games
640
Steps covered
46 (step 0–96)
Skill
Games
Causal Inference
96
Logical Deduction
115
Mathematical Reasoning
120
Optimization
97
Pattern Recognition
100
Spatial Reasoning
112
Layout
manifest.json… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-30b-plateau-kl0-spare-games-envs.nba-games
NBA Games Data
This data is an updated version of the original NBA
Games by Nathan Lauga.
Data source
Code
Updated to: 2025-02-13
The dataset retains the original format and includes the following files:
games.csv – Summary of NBA games, including scores and team details.
games_details.csv – Detailed player statistics for each game.
players.csv – Player information.
ranking.csv – Daily NBA team rankings.
teams.csv – List of all NBA teams.
qwen3-30b-0617-plateau-6skill-nomem-spare-games-envs
qwen3-30B-A3B-Instruct-0617-plateau-6skill-nomem — generated environments
Environments generated by the SPARE proposer during training run
z1wnr4j9 (qwen3-30B-A3B-Instruct-0617-plateau-6skill-nomem), recovered from the spare-viz durable cache.
The run's scratch directory no longer exists; this dataset is the surviving copy.
Games
3746
Steps covered
170 (step 0–358)
With recovered skill
3692
With hint
0
Actor / proposer model… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-30b-0617-plateau-6skill-nomem-spare-games-envs.c2c-ai-vs-ai
C2C: Cooperate to Compete — AI vs AI Games
This dataset contains 972 fully-logged AI vs AI games from the Cooperate to Compete (C2C) benchmark — a long-horizon, mixed-motive multi-agent negotiation environment based on a four-player conquest game with private regional objectives, fog of war, and non-binding cheap-talk negotiation.
Project page: https://negotiationgame.io/c2c/
Paper: https://arxiv.org/abs/2604.25088
Play against AI agents: https://negotiationgame.io
Github:… See the full description on the dataset page: https://huggingface.co/datasets/negotiation-games/c2c-ai-vs-ai.SPADE-Environments-Qwen3-30B-Games
SPADE generated environments: games
Paper | Code | All artifacts
Executable game environments written by the SPADE Environment Designer during the paper's 30B games self-play run. One Python file per environment; manifest.json records the generation checkpoint, training step, skill, and difficulty of each.
Environments
3310
Training steps covered
113 (step 0 to 396)
With skill label
3119
Designer / agent model
Qwen/Qwen3-30B-A3B-Instruct-2507… See the full description on the dataset page: https://huggingface.co/datasets/spade-rl/SPADE-Environments-Qwen3-30B-Games.qwen3-8b-0708-games-blend-bothink-kl005-spare-games-envs
qwen3-8B-0708-games-blend-bothink-kl005 — generated environments
Environments generated by the SPARE proposer during training run
0n8pbtct (qwen3-8B-0708-games-blend-bothink-kl005), recovered from the spare-viz durable cache.
The run's scratch directory no longer exists; this dataset is the surviving copy.
Games
3702
Steps covered
106 (step 0–388)
With recovered skill
3497
With hint
0
Actor / proposer model
/scratch/spare-workspace/Qwen3-8B
WandB segments… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-8b-0708-games-blend-bothink-kl005-spare-games-envs.chess_gamesDataset descriptions:
lichess_6gb: 6GB of 16 million games from lichess's database. 16492151 games, 6486463314 chars. No elo filtering performed. Comprised of games from lichess 2016-06 and 2017-05.
lichess_9gb: 9GB of games from lichess's database. No elo filtering performed. Comprised of games from lichess 2017-07 and 2017-08.
lichess_100mb: 100MB of 300k games from lichess's database. Comprised of games from lichess 2016-01. This is used to train linear probes on a separate dataset from… See the full description on the dataset page: https://huggingface.co/datasets/adamkarvonen/chess_games.qwen3-30b-0705c-glory-r8-premerge-spare-games-envs
qwen3-30B-A3B-Instruct-0705c-glory-r8-premerge — generated environments
Environments generated by the SPARE proposer during training run
5hvg1dna (qwen3-30B-A3B-Instruct-0705c-glory-r8-premerge), recovered from the spare-viz durable cache.
The run's scratch directory no longer exists; this dataset is the surviving copy.
Games
4023
Steps covered
105 (step 0–392)
With recovered skill
4023
With hint
0
Actor / proposer model… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-30b-0705c-glory-r8-premerge-spare-games-envs.qwen3-30b-plateau-kl005-spare-games-envs
qwen3-30B-A3B-Instruct plateau-6skill KL=0.05 — generated environments
Environments generated during the 30B plateau KL=0.05 run (main segment 20260718_103657), recovered from the run's surviving on-disk game cache.
Games
480
Steps covered
24 (step 0–76)
Skill
Games
Causal Inference
81
Logical Deduction
80
Mathematical Reasoning
81
Optimization
79
Pattern Recognition
80
Spatial Reasoning
79
Layout
manifest.json… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-30b-plateau-kl005-spare-games-envs.tournament-chess-games
Lichess Broadcasts Dataset
Dataset Description
931,021 chess games from chess tournaments tracked using Lichess Broadcasts.
Lichess Broadcasts show live games as they unfold with new moves arriving in real time. They are built to connect to the live-updating PGN file produced by DGT boards but can work with other sources as well.
Broadcasts are organized in "tournaments" and "rounds."
Dataset Sample
{
'Event': 'FIDE World Championship Match 2021',
'Site':… See the full description on the dataset page: https://huggingface.co/datasets/Lichess/tournament-chess-games.chess960-chess-games
[!CAUTION]
This dataset is still a work in progress and some breaking changes might occur.
Note
The FEN column has 961 unique values instead of the expected 960, because some rematches were recorded with invalid castling rights in their starting FEN in November 2023.
gamesdb_public_dataset
GAMESDB / PUBLIC DATASETS
CHECK "FILES AND VERSIONS" FOR THE DATASETS
This is intended to hold all the datasets that GamesDB will be using from a variety of sources.
You are allowed to use these datasets freely! They are available publically.
NOTE: depending at what time your reading this, dataset_steam1.csv and dataset_playstore1.csv are likely unavailible and will be released by the end of this week here. You can still download them from their original source.
gen-games-v7-video-pilot1qwen3-30b-0617-plateau-6skill-nocorpus-spare-games-envs
qwen3-30B-A3B-Instruct-0617-plateau-6skill-nocorpus — generated environments
Environments generated by the SPARE proposer during training run
5uui7594 (qwen3-30B-A3B-Instruct-0617-plateau-6skill-nocorpus), recovered from the spare-viz durable cache.
The run's scratch directory no longer exists; this dataset is the surviving copy.
Games
28620
Steps covered
223 (step 0–395)
With recovered skill
28237
With hint
0
Actor / proposer model… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-30b-0617-plateau-6skill-nocorpus-spare-games-envs.qwen3-8b-0627-plateau-6skill-thinking-proposer-spare-games-envs
qwen3-8B-0627-plateau-6skill-thinking-proposer — generated environments
Environments generated by the SPARE proposer during training run
sh04swu4 (qwen3-8B-0627-plateau-6skill-thinking-proposer), recovered from the spare-viz durable cache.
The run's scratch directory no longer exists; this dataset is the surviving copy.
Games
1411
Steps covered
73 (step 0–149)
With recovered skill
1312
With hint
0
Actor / proposer model
/scratch/spare-workspace/Qwen3-8B… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/qwen3-8b-0627-plateau-6skill-thinking-proposer-spare-games-envs.PS-Games-Datasetgen-games-v9-video-pilot1lc0_games
[!WARNING]
Due to a parsing error, the castling moves in the chess960 games are encoded as (king starting square) -> (king target square) and are not compatible with python-chess.
A workaround is shown in the Use section.
LC0 Games
This dataset contains 170m games played by Leela Chess Zero against itself. The moves are written in uci format.
The games were sourced directly from the lc0 official data.
This dataset contain some chess960 games, for which the start FEN is… See the full description on the dataset page: https://huggingface.co/datasets/groloch/lc0_games.