datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
clash-royale-tv-replays
Clash Royale TV Replays
Frame-by-frame gameplay recordings (~10 fps) from Clash Royale's TV Royale, covering all 31 arenas. Automated recording using tools from our github repository.
Structure
arena_{XX}/{replay_uuid}/
├── frames.parquet # Frame data
└── preview.jpg # First frame thumbnail
Parquet Schema:
frame_id (int64): Frame number
image (Image): PNG bytes
hash (string): MD5 for deduplication
Usage
from huggingface_hub import hf_hub_download… See the full description on the dataset page: https://huggingface.co/datasets/chrisrca/clash-royale-tv-replays.clash-royale-tv-replays
Clash Royale TV Replays
Frame-by-frame gameplay recordings (~10 fps) from Clash Royale's TV Royale, covering all 31 arenas. Automated recording using tools from our github repository.
Structure
arena_{XX}/{replay_uuid}/
├── frames.parquet # Frame data
└── preview.jpg # First frame thumbnail
Parquet Schema:
frame_id (int64): Frame number
image (Image): PNG bytes
hash (string): MD5 for deduplication
Usage
from huggingface_hub import hf_hub_download… See the full description on the dataset page: https://huggingface.co/datasets/Chutdedededede/clash-royale-tv-replays.melee-ranked-replays
Melee Ranked Replays
Anonymized Slippi ranked replays (platinum+) from Super Smash Bros. Melee,
sharded by character and rank pair. Built for behavior-cloning and other
replay-driven ML work on Melee — notably MIMIC.
Contents
Raw .slp files grouped into tarballs by (character, rank_pair, source_archive),
organized into per-character folders:
{CHAR}/
{CHAR}_{rank_pair}_a{N}.tar.gz
metadata/
metadata_a{N}.json
Characters (25): BOWSER, CPTFALCON, DK, DOC, FALCO… See the full description on the dataset page: https://huggingface.co/datasets/erickfm/melee-ranked-replays.pokemon-showdown-replays
Pokémon Showdown Replay Datasets
Multiple datasets of Pokémon Showdown public replays pulled using the Pokémon Showdown API. Publishing them here to provide them to the community as a starting point for future projects.
These datasets are used in Showdown Shower, an university project of mine.
The dates of the replays range from 2005 to 2026.
Currently (as of 17th September, 2026), this is the number of replays available in each dataset. The number may be lower than the one… See the full description on the dataset page: https://huggingface.co/datasets/HolidayOugi/pokemon-showdown-replays.IL_Replay
IL_Replay
An anonymized battle replay dataset for imitation learning and offline AI research: 252,238 replays and 17,836,160 actions. The replays and actions configurations expose the two related tables separately. All records are in the train split.
本目录合并了 252,238 场回放和 17,836,160 条动作记录。
目录
replays/part-*.parquet：对局元数据与完整 payload_json，用于 Firstlight_CR 的训练缓存生成和采集回放功能。
actions/part-*.parquet：展开的动作表，通过新的 replay_tag 与回放表关联。完整动作也保存在回放 JSON 中。… See the full description on the dataset page: https://huggingface.co/datasets/VanguardX101/IL_Replay.dclm-replay.seq-4096.tokens-32B2^35 tokens of replay data from DCLM-baseline, concatenated into 2^23 sequences of 4096 tokens each with <|endoftext|> separators.
replay-mix-10Bmetamon-raw-replays
Metamon Raw Replay Dataset
This dataset maintains the up-to-date supply of Pokémon Showdown replays used by the Metamon project. Battles are publicly available from the Showdown replay API. We provide them here to create a convenient starting point for building new parsed-replay verisons of the RL dataset, and maybe to save Showdown some download requests. Usernames have been replaced by a unique randomly generated pseudonym that is consistent across battles.
Versions… See the full description on the dataset page: https://huggingface.co/datasets/jakegrigsby/metamon-raw-replays.pokemon-showdown-replays-merged
Pokemon Showdown Replays - Merged Dataset
A merged dataset of Pokemon Showdown battle replays from multiple sources.
Statistics
Total Replays: 29,057,184
Sources:
sethkarten: 3,909,792 replays
metamon: 1,978,010 replays
holidayougi: 23,169,382 replays
Top Formats
Format
Count
[Gen 9] OU
3,200,571
[Gen 6] OU
2,977,289
[Gen 7] OU
2,578,070
[Gen 7] RANDOMBATTLE
2,005,704
[Gen 6] RANDOMBATTLE
1,598,899
[Gen 9] VGC 2025
836,958
[Gen 9]… See the full description on the dataset page: https://huggingface.co/datasets/milkkarten/pokemon-showdown-replays-merged.metamon-parsed-replays
Metamon Replay Dataset
Pokémon Showdown replay files parsed (or "reconstructed") into RL trajectories by Metamon (arXiv Appendix D)
Quick Start
The easiest way to use the replay dataset is through metamon's dataloader:
import metamon
from metamon.interface import get_observation_space, get_reward_function, get_action_space
from metamon.data importParsedReplayDataset
# see the metamon README for more on observations, actions, and rewards.
human_dset =… See the full description on the dataset page: https://huggingface.co/datasets/jakegrigsby/metamon-parsed-replays.tulu3-sft-replay-othello-500k
Tulu-3 SFT replay subset (Llama-3 chat)
A randomly-sampled, token-sized subset of allenai/tulu-3-sft-mixture, for use as replay data when fine-tuning on a narrow board-game task (Othello / Snake-Othello), to preserve general instruction-following.
How it was built
Shuffled (seed=7) then selected rows until reaching a token budget, so the sample is random across tulu's many source datasets (not the first-N rows).
Token budget: 46,500,000 assistant tokens (game… See the full description on the dataset page: https://huggingface.co/datasets/cfierro/tulu3-sft-replay-othello-500k.uniref50-replay-mix-v1
uniref50-replay-mix-v1
Stage-1 continued-pretraining corpus for
eshmun-vocab: protein sequences (UniRef50) mixed with
a general/biomedical/math/code text replay slice, so a vocab-extended LLM
(e.g. khairi/qwen3-0.6b-protein-vocab-v0) learns protein-sequence
statistics without catastrophically forgetting its pretrained language
ability. Design rationale and target ratios: see docs/pretrain-dataset-mix.md
in the eshmun-vocab repo.
Objective: plain next-token prediction. Every row… See the full description on the dataset page: https://huggingface.co/datasets/khairi/uniref50-replay-mix-v1.MLLM-CL-ReplayData
MLLM-CL: Continual Learning for Multimodal Large Language Models
This is the official dataset repository of MLLM-CL and MR-LoRA. MLLM-CL is a novel benchmark encompassing domain and ability continual learning, where the former focuses on independently and identically distributed (IID) evaluation across evolving mainstream domains, whereas the latter evaluates on non-IID scenarios with emerging model ability. MR-LoRA prevents catastrophic interference through parameter isolation and… See the full description on the dataset page: https://huggingface.co/datasets/MLLM-CL/MLLM-CL-ReplayData.worldforge-go2-dimos-replay-world-pairs
WorldForge Go2 DimOS Replay World Pairs
This dataset is a compact, derived world-model dataset built from public
dimensionalOS/dimos Unitree Go2 replay assets.
Companion benchmark: go2-air-controlbench-v1
provides measured command-to-outcome trials on a real Go2 (commands, no images). This
dataset provides the robot-POV image pairs (images, no commands). Together they cover the
visual and control halves of the WorldForge score workflow.
New — expanded config:… See the full description on the dataset page: https://huggingface.co/datasets/espejelomar/worldforge-go2-dimos-replay-world-pairs.deepseek-v4-flash-swebench-replay
deepseek-v4-flash-swebench-replay
中文
这是一个 DeepSeek V4 Flash 在 SWE-bench 上的 agentic replay 数据集仓库。
它的目标是让使用者不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，就可以直接查看和重放模型的多轮推理与工具调用轨迹。
当前包含的数据
verified_agentic
lite_agentic
当前不包含的数据
单轮 single-turn trace
verified_mini_agentic（当前本地仅完成 31/50，因此不纳入首版）
分数汇总
verified_agentic: 354 / 500, Acc/Pass@1 = 70.8
lite_agentic: 182 / 300, Acc/Pass@1 = 60.67
数据来源
这些轨迹由… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/deepseek-v4-flash-swebench-replay.dclm-replay.seq-4096.n-262144-llama32dclm-replay.seq-1024.n-1048576-llama32tabrepairbench-replayable-corruption-lineage
TabRepairBench: Replayable Corruption Lineage
This is a finite, wholly authored synthetic reference dataset for auditing
tabular corruption lineage. Public Lineage v1 contains 72 independent groups
and 6,912 clean/corrupt cell pairs across three authored structural generator
families, six corruption schedules, two severities, and two audit partitions.
It is intentionally not presented as real-world data. It makes no claim of
real-data representativeness, causal identification… See the full description on the dataset page: https://huggingface.co/datasets/haidang2405/tabrepairbench-replayable-corruption-lineage.OpenMathReasoning_len8k_0.2_replay2kdclm-replay.seq-4096.n-262144-olmo3
dclm-replay.seq-4096.n-262144-olmo3
OLMo 3-tokenized CPT-replay sequences for prestar, the OLMo analogue of JackHsieh/dclm-replay.seq-4096.tokens-32B (Qwen3).
Source: mlfoundations/dclm-baseline-1.0, pin global-shard_01_of_10/local-shard_0_of_10/*.jsonl.zst.
Tokenizer: allenai/Olmo-3-1025-7B; EOD token id 100257 (<|endoftext|>).
262,144 sequences of exactly 4096 tokens each (docs concatenated and packed; EOD-separated).
Same builder/pin as the Qwen3 replay — corpus is the same… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/dclm-replay.seq-4096.n-262144-olmo3.TTS_replay-15_ls960-testarch-selection-replay
Design selection replay: two trained campaigns and every selector tried
The records needed to score a design selector offline, in the shape of xRouteBench: every candidate design was trained once and the outcome kept, so a rule that picks which unexecuted design gets the GPU is evaluated by replay against the same rows as every previous rule, with no GPU and no model call. Two campaigns, 24 in-sample and 15 held-out designs, each row carrying what a selector may see (task… See the full description on the dataset page: https://huggingface.co/datasets/neurarch-ai/arch-selection-replay.TTS_replay-14_ls960-testOpenMathReasoning_len8k_0.6_self_replay2k10gen9-random-doubles-replays
Gen 9 Random Doubles replays
Every gen9randomdoublesbattle replay Pokemon Showdown still lists, plus the
ones it has since delisted, deduplicated on replay id.
229,171 battles, 2023-06-26 to 2026-08-26. 73.2% carry a rating: 61,673 at
1400 or better, 43,756 at 1500, 27,728 at 1600, 14,506 at 1700, 5,555 at 1800.
Why this exists
Neither available source is complete on its own.
HolidayOugi/pokemon-showdown-replays
holds 203,654 distinct battles for this format and… See the full description on the dataset page: https://huggingface.co/datasets/armaanp/gen9-random-doubles-replays.TTS_replay_ls960-testhedgehog-schema-replay
hedgehog-schema-replay
Hedgehog — schema replay round (anti-forgetting).
Contents
train.jsonl (1360 rows)
validation.jsonl (160 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Original content for the Hedgehog extraction model (Michael Anthony Falabella).
GLM51-swebench-replay
GLM51-swebench-replay
中文
这是一个 GLM-5.1 在 SWE-bench 上的 agentic replay 数据集仓库。目标是让使用者不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，就可以直接查看和重放模型的多轮推理、工具调用和最终 patch。
数据来源
轨迹使用 EvalScope 收集，benchmark 使用 EvalScope 中的 official SWE-bench agentic 数据集：
swe_bench_verified_agentic
swe_bench_lite_agentic
运行时使用 SWE-bench 官方容器镜像，收集形态为 agentic + toolcall。模型调用使用 GLM-5.1 的 OpenAI-compatible 接口。
分数汇总
verified_agentic: 363 / 500, Acc/Pass@1 = 72.6… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/GLM51-swebench-replay.TTS_replay_2_ls960-testgenerals_io_replays
⚔️ Generals.io High-Rank Replay Dataset 🌟
Overview
This dataset contains a curated collection of 1v1 game replays from the online strategy game generals.io, specifically designed for training high-level reinforcement learning agents 🤖.
🏆 High-Quality Matches: Includes games where at least one participant had a star rating of 70 or higher 📈, ensuring a baseline of quality and strategic depth.
✅ Clean Data: Carefully filtered to remove outliers, games with AFK players… See the full description on the dataset page: https://huggingface.co/datasets/strakammm/generals_io_replays.