datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
minimax_h3_1k
MiniMax H3 - 1K
I generated a dataset to test the knowledge scope and capabilities of MiniMax H3.
Samples are of various aspect sizes, and cover a wide range of media types and themes.
The videos are 768 base resolution (~0.6 MP).
They were generated with minimax_h3_fl2va_pruned_int8_convrot.safetensors with 30 steps.
A full video of this dataset can be viewed here https://youtu.be/akkwj9d943Y
minimax-h3-soup
MiniMax H3 Soup
Reproducibility archive for a local ComfyUI MiniMax H3 Ref2V benchmark on an RTX 3090.
What is included
Original benchmark workflow graph (source_prompt.json), manifest, and result table.
Every one-second MP4 from the original C1-C11 benchmark grid and its Euler
repeat sweep. The separate
duration experiments are intentionally not included.
Labeled C1-C11 visual contact sheets, Ref2VA stock-control sheets, and a
static render-time summary chart.… See the full description on the dataset page: https://huggingface.co/datasets/badincite/minimax-h3-soup.SynLogic
SynLogic Dataset
SynLogic is a comprehensive synthetic logical reasoning dataset designed to enhance logical reasoning capabilities in Large Language Models (LLMs) through reinforcement learning with verifiable rewards.
🐙 GitHub Repo: https://github.com/MiniMax-AI/SynLogic
📜 Paper (arXiv): https://arxiv.org/abs/2505.19641
Dataset Description
SynLogic contains 35 diverse logical reasoning tasks with automatic verification capabilities, making it ideal for… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/SynLogic.MiniMax-M2.1-Mixture-of-Thoughts
MiniMax-M2.1 Mixture of Thoughts
This dataset contains responses generated by MiniMax-M2.1 for user questions from the open-r1/Mixture-of-Thoughts dataset.
Dataset Description
The dataset captures both the extended thinking process and final answers from MiniMax-M2.1, with reasoning wrapped in <think> tags for easy separation.
Metric
Value
Examples
349,317
Total Tokens
4,052,592,552
Avg Tokens/Example
11,601
Source Dataset
Name:… See the full description on the dataset page: https://huggingface.co/datasets/PursuitOfDataScience/MiniMax-M2.1-Mixture-of-Thoughts.TTS-Multilingual-Test-Set
Overview
To assess the multilingual zero-shot voice cloning capabilities of TTS models, we have constructed a test set encompassing 24 languages. This dataset provides both audio samples for voice cloning and corresponding test texts.
Specifically, the test set for each language includes:
100 distinct test sentences.
Audio samples from two speakers (one male and one female) carefully selected from the Mozilla Common Voice (MCV) dataset, intended for voice cloning.
Researchers can… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/TTS-Multilingual-Test-Set.jav-minimax-h3minimax-music-3-datasetMiniMax Music 3.0 Dataset
A large synthetic MiniMax Music research dataset by Angelware Research
8,681 tracks generated with MiniMax Music 3.0 for audio analysis, benchmarking, provenance research, and AI-music detection.
At a glance
Generated with MiniMax Music 3.0. Audio is preserved exactly as received, including embedded AIGC provenance tags where present.
Collection statistic
Value
Tracks
8,681
Total duration… See the full description on the dataset page: https://huggingface.co/datasets/AngelSoftware/minimax-music-3-dataset.minimax_h3_avatar_500
Watch the full 500-video showcase on YouTube
MiniMax H3 Avatar 500
An image-to-video dataset pairing reference avatar images with detailed generation prompts and generated avatar videos. This release contains 500 curated examples in both a browsable raw layout and a typed Hugging Face dataset.
Version 1.0 · Released August 14, 2026
Dataset contents
Each example contains:
A 1024 × 1024 reference avatar image
A detailed English generation prompt
A generated 640 ×… See the full description on the dataset page: https://huggingface.co/datasets/oakmindai/minimax_h3_avatar_500.minimax-m3-claude-code-tracesThis dataset was generated using teich by TeichAI
Prepare these datasets for supervised fine-tuning in just a few lines of code — see the Conversion section below.
Minimax M3 Claude Code Traces
This directory contains raw agent trace files generated by teich.
All assistant responses were generated by minimax/minimax-m3.
JSONL files: 31
Format
Each file is newline-delimited JSON representing a single captured agent session.
The trace schema is designed for… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/minimax-m3-claude-code-traces.Minimax-H3
Install Minimax H3 in ComfyUI (NVFP4/BF16/FP8/GGUF):
https://www.stablediffusiontutorials.com/2026/08/minimax-h3.html
SWE-smith-rs-minimax-m2.5-trajectories
Trajectories Dataset
Top-level fields:
messages
instance_id
resolved
model
traj_id
patch
Generated at: 2026-02-27 00:04:28Z
Rows: 5251
Shards: 21
Skipped runs (missing/corrupt trajectory): 60
swerebench-filtered-openhands-minimax-m2_5-corrected-completionsrole-play-bench
Role-play Benchmark
A comprehensive benchmark for evaluating Role-play Agents in Chinese and English scenarios.
Dataset Summary
Role-play Benchmark is designed to evaluate Role-play Agents' ability to deliver immersive role-play experiences through Situated Reenactment. Unlike traditional benchmarks with verifiable answers, Role-play is fundamentally non-verifiable, e.g., there's no single "correct" response when a tsundere character is asked "Do you like me?".… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/role-play-bench.OctoCodingBench
OctoCodingBench: Instruction-Following Benchmark for Coding Agents
English | 中文
🌟 Overview
OctoCodingBench benchmarks scaffold-aware instruction following in repository-grounded agentic coding.
Why OctoCodingBench?
Existing benchmarks (SWE-bench, etc.) focus on task completion — whether the agent produces correct code. However, they miss a critical dimension: does the agent follow the rules while solving the task?
In real-world agentic coding, agents must… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/OctoCodingBench.minimax-h3-qkv-ksnr-20260917minimax-m2Minimax_H3_Parasyte_fasth3_6step_testPurpose of this Minimax H3 Turbo lora test is to find the best quality settings for low resolution generation for my old trusted RTX 3060 12Gb. All clips were created with 416x768 resolution.
Workflow used was modified PlagueKind's V7 with Mikey's Wildcard Processor, LLM enhancement node (Minimax-H3-Prompt-Rewriter) and Qwen3VL 4b used as an LLM and text encoder via ClipProj node.
Checkpoint and other settings in the workflow used are by default, or near what Plaguekind has recommended. If… See the full description on the dataset page: https://huggingface.co/datasets/Turpo/Minimax_H3_Parasyte_fasth3_6step_test.minimax-m3-deepsearchqa-skill-eval
MiniMax M3 DeepSearchQA Skill Eval
Evaluates minimax/minimax-m3 on google/deepsearchqa using a Pi agent, You.com MCP tools, and a research skill optimized for this harness, model, and tool surface.
MiniMax M3 Medium Reasoning with the You.com research skill reached 74.85% adjusted F1 on DeepSearchQA, above the paper's GPT-5 High Reasoning F1 result. Public artifacts are available for inspection and reproduction.
Links
GitHub:… See the full description on the dataset page: https://huggingface.co/datasets/youdotcom/minimax-m3-deepsearchqa-skill-eval.VIBE
VIBE: Visual & Interactive Benchmark for Execution in Application Development
[English] | 中文
🌟 Overview
VIBE (Visual & Interactive Benchmark for Execution) sets a new standard for evaluating Large Language Models (LLMs) in full-stack software engineering. Moving beyond recent benchmarks that rely on static screenshots or rigid workflow snapshots to assess application development, VIBE pioneers the Agent-as-a-Verifier (AaaV) paradigm to assess the true "0-to-1" capability… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/VIBE.minimax-m2.7-agent
Agentic Training Traces
This directory contains raw agent trace files generated by agentic-datagen.
All assistant responses were generated by minimax/minimax-m2.7.
Trace files: 20
Training-ready tools
Use this tools payload when rendering converted examples through your training chat template.
The same structure is emitted on each converted example as the tools field.
[
{
"type": "function",
"function": {
"name": "bash",
"parameters": {… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/minimax-m2.7-agent.open-qwen-music-dataset-c-minimax-music3
Open-Qwen-Music Dataset C + MiniMax-Music3
This release contains 62,417 audio records (2,958.571 hours).
It combines an access-screened subset of Dataset C with publisher-confirmed
MiniMax-Music3 outputs. It also provides 631
Cambridge-MT metadata-only records.
Usage restriction
This package may be used only for non-commercial academic research. Commercial use
is prohibited. This is a research-only distribution, not an open-source license under
the Open… See the full description on the dataset page: https://huggingface.co/datasets/david-miller-45678/open-qwen-music-dataset-c-minimax-music3.nemotron-gym-instruction-following-structured-minimax-m27-131k-tracesMinimax-H3
MiniMax-H3 Human–Scene Interaction Dataset
894 paired samples of video + text prompt + 3D human motion (SMPL) for training
joint video–motion generation models. Each clip shows a single person performing a
language-specified human–scene interaction (walking to a bench and sitting down,
leaning against a stone, etc.) under a locked-off static camera, with the full body
visible in frame at all times.
Videos were generated with MiniMax-H3 (first-frame image + prompt), and per-frame… See the full description on the dataset page: https://huggingface.co/datasets/Fyantu/Minimax-H3.minimax-music3-rvq-reverse-distillation
MiniMax Music 3 RVQ Reverse-Distillation Traces
Research traces generated from MiniMax Music 3. Each ZIP contains generated
audio, sampled RVQ codes, teacher sampling logits, conditioning embeddings,
Flow-VAE latents, and source prompt metadata. See campaign-config.json for
generation settings and indexes/ for per-commit manifests.
Dataset statistics
As of 2026-08-16, the uploaded snapshot contains:
2,972 generated songs
91.84 hours of actual generated audio
1… See the full description on the dataset page: https://huggingface.co/datasets/bghira/minimax-music3-rvq-reverse-distillation.minimax-music3-rvq-distill-corpus-8k
MiniMax Music3 self-distillation corpus — 11,847 tracks / ~193 h
Paired (audio, RVQ codes, teacher top-50 distributions, DAV latents) traces generated with the official
MiniMax Music3 pipeline (diffusers ModularPipeline), built to train/improve the community audio->codes
RVQ encoder. Fine-tuning SimpleTuner/open-rvq-encoder-minimax-music3-41m-v1 on this corpus pooled with its
original data improves every holdout metric — see… See the full description on the dataset page: https://huggingface.co/datasets/Mothersuperior/minimax-music3-rvq-distill-corpus-8k.minimax-h3-nvfp4-data
MiniMax-H3 NVFP4 Data
Large data for the NVFP4 static sparse attention experiment (2026-09-11).
Project entry point: H3 Attention Lab.
The project README contains result tables, audit limitations, source code, small evidence files, and public download instructions.
This revision contains 84 data files (752,104,680 bytes):
70 generated videos: experiment/outputs/*/*/quality.mp4
3 comparison videos: experiment/comparisons/*__e2e_comparison.mp4
10 reference videos:… See the full description on the dataset page: https://huggingface.co/datasets/Zer0-Sky/minimax-h3-nvfp4-data.minimax-h3-video-prompts
MiniMax H3 Video Prompts
A small, curated collection of 50 structured prompts for text-to-video and image-to-video workflows. It covers cinematic scenes, characters, animation, nature, architecture, product shots, food, social video, and fantasy environments.
Use the prompts to create video
Copy a prompt from the dataset, adapt it to your idea, then generate the finished video online.
Create an AI video with MiniMax3.org →
Dataset details… See the full description on the dataset page: https://huggingface.co/datasets/jayseanbrambila/minimax-h3-video-prompts.minimax-h3-prompt-dataset
MiniMax H3 Structured Video Prompts
A community dataset of 300 structured MiniMax H3 video prompts for text-to-video and image-to-video workflows, covering product advertising, e-commerce, fashion, beauty, food, automotive, SaaS, social media, camera motion, lighting, and commercial-use scenarios.
Each record separates reusable prompting elements such as subject, action, environment, camera motion, lighting, visual style, workflow, use case, and commercial intent.… See the full description on the dataset page: https://huggingface.co/datasets/jayseanbrambila/minimax-h3-prompt-dataset.minimax-m3-tiny-cpu-repro-v1
minimax-m3 complete native tiny random CPU fixture
Complete untrained MiniMaxM3SparseForConditionalGeneration checkpoint with an untied full LM head,
a real 272-entry byte tokenizer and every native state tensor. Architecture lineage:
MiniMaxAI/MiniMax-M3@f0e1c1e04d40177e4673a22097036854f536e9c0.
No upstream weights, training data, paid GPU or cloud compute were used.
Complete native image/text wrapper with real shrunk Conv3D vision, nonempty 3D RoPE, patch-merge projector and… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/minimax-m3-tiny-cpu-repro-v1.minimax-m2-tiny-cpu-repro-v1
minimax-m2 complete native tiny random CPU fixture
Complete untrained MiniMaxM2ForCausalLM checkpoint with an untied full LM head,
a real 272-entry byte tokenizer and every native state tensor. Architecture lineage:
MiniMaxAI/MiniMax-M2.7@d494266a4affc0d2995ba1fa35c8481cbd84294b.
No upstream weights, training data, paid GPU or cloud compute were used.
Complete native text causal LM: sigmoid/top-k MoE routing with correction bias, per-layer flattened Q/K RMSNorm and half-head… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/minimax-m2-tiny-cpu-repro-v1.