datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.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.role-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.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-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.OctoBench
OctoBench: Instruction-Following Benchmark for Coding Agents
English | 中文
📰 News
[2026-01-16] 📄 Paper released on arXiv: OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding[2026-01] 🎉 Dataset released on Hugging Face
🌟 Overview
OctoBench is an extended version of OctoCodingBench, expanded from the original 72 manually annotated instances to 217 instances using AI-assisted Augmentation.
📦… See the full description on the dataset page: https://huggingface.co/datasets/MiniMaxAI/OctoBench.MiniMax-M3-150k-Mixed
m3-alldomains-verified-107k
Verified distillation traces generated with faststill v0.0.1 — a pipeline that generates (prompt, reasoning, output) triplets from any OpenAI-compatible chat-completions endpoint and deterministically verifies every row before keeping it. A row is verified=true only when a machine check (executed unit tests, exact / normalized answer compare) confirmed it, so wrong labels are filtered out instead of poisoning a student model.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/empero-ai/MiniMax-M3-150k-Mixed.reasoning-sft-minimax-microsoft-orca-agentinstruct-1M-v1
MiniMax-M2.5 Reasoning SFT (Orca AgentInstruct 1M v1)
Reasoning SFT dataset generated by MiniMaxAI/MiniMax-M2.5 on prompts from the Stratified K-Means Diverse Instruction-Following 100K-1M dataset (Orca AgentInstruct subset).
Format
Each row has three columns:
input — list of dicts [{"role": "...", "content": "..."}, ...] (conversation turns)
response — model-generated response with <think> reasoning block
source — task category (creative_content, text_modification, rc… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-minimax-microsoft-orca-agentinstruct-1M-v1.swebench-minimax-traces
swebench-minimax-traces
Complete HTTP-level agentic traces from running swebench benchmark tasks through an instrumented reverse proxy.
Each trace captures full request/response pairs including system prompts, user messages, assistant responses, tool calls and results, and token usage metadata.
Stats
Total sessions: 500
Multi-turn sessions (2+ LLM calls): 357
Total records: 22424
Total LLM requests: 11249
Format
Raw JSONL traces from the instrumented proxy.… See the full description on the dataset page: https://huggingface.co/datasets/sammshen/swebench-minimax-traces.Sonnet-Opus-4.5-4.6-Gemini-3.0-3.1-Pro-GPT-5-5.1-5.2-GLM-4.7-MiniMax-M2.1-DeepSeek-V3.2-High
Distill
This is a multi-source curated instruction and reasoning dataset specifically for training and distilling large language models (LLMs) to exhibit advanced Chain-of-Thought (CoT), Agentic, Mathematical and Coding capabilities. It aggregates high-quality outputs from frontier models into messages ChatML format.
Dataset Structure
The dataset contains a total of 70.2K examples, split into three subsets based on the presence of visible reasoning… See the full description on the dataset page: https://huggingface.co/datasets/VINAY-UMRETHE/Sonnet-Opus-4.5-4.6-Gemini-3.0-3.1-Pro-GPT-5-5.1-5.2-GLM-4.7-MiniMax-M2.1-DeepSeek-V3.2-High.MiniMax-M3-70k-CodeMath
m3-math-code-verified-100k
Verified distillation traces generated with faststill v0.0.1 — a pipeline that generates (prompt, reasoning, output) triplets from any OpenAI-compatible chat-completions endpoint and deterministically verifies every row before keeping it. A row is verified=true only when a machine check (executed unit tests, exact / normalized answer compare) confirmed it, so wrong labels are filtered out instead of poisoning a student model.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/empero-ai/MiniMax-M3-70k-CodeMath.Medical-Reasoning-SFT-MiniMax-M2.1
Medical-Reasoning-SFT-MiniMax-M2.1
A large-scale medical reasoning dataset generated using MiniMaxAI/MiniMax-M2.1, containing over 204,000 samples with detailed chain-of-thought reasoning for medical and healthcare questions.
Dataset Overview
Metric
Value
Model
MiniMaxAI/MiniMax-M2.1
Total Samples
204,773
Samples with Reasoning
204,773 (100%)
Estimated Tokens
~621 Million
Content Tokens
~344 Million
Reasoning Tokens
~277 Million
Language
English… See the full description on the dataset page: https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-MiniMax-M2.1.intercode-minimax-traces
intercode-minimax-traces
Complete HTTP-level agentic traces from running intercode benchmark tasks through an instrumented reverse proxy.
Each trace captures full request/response pairs including system prompts, user messages, assistant responses, tool calls and results, and token usage metadata.
Stats
Total sessions: 338
Multi-turn sessions (2+ LLM calls): 299
Total records: 5838
Total LLM requests: 2919
Format
Raw JSONL traces from the instrumented proxy.… See the full description on the dataset page: https://huggingface.co/datasets/sammshen/intercode-minimax-traces.reasoning-sft-minimax-stratified-kmeans-diverse-reasoning-842K-only
MiniMax-M2.5 Reasoning SFT (Stratified K-Means Diverse Reasoning 1M)
Reasoning SFT dataset generated by MiniMaxAI/MiniMax-M2.5 on prompts from the Stratified K-Means Diverse Reasoning 100K-1M dataset.
Format
Each row has three columns:
input — list of dicts [{"role": "...", "content": "..."}, ...] (conversation turns)
response — model-generated response with <think> reasoning block
source — task category (math, code, science, chat, safety)
Generation
Model:… See the full description on the dataset page: https://huggingface.co/datasets/AmanPriyanshu/reasoning-sft-minimax-stratified-kmeans-diverse-reasoning-842K-only.minimax-m3-150k-mixed
m3-alldomains-verified-107k
Verified distillation traces generated with faststill v0.0.1 — a pipeline that generates (prompt, reasoning, output) triplets from any OpenAI-compatible chat-completions endpoint and deterministically verifies every row before keeping it. A row is verified=true only when a machine check (executed unit tests, exact / normalized answer compare) confirmed it, so wrong labels are filtered out instead of poisoning a student model.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/ansulev/minimax-m3-150k-mixed.minimax-m2.1-reap-observations
[!TIP]
Support this work: donate.sybilsolutions.ai
REAP surfaces: GLM | MiniMax | Qwen | Gemma | Paper | Code | PR17 | Cerebras Collection
MiniMax-M2.1 REAP Stress Test Observations
Comprehensive stress test results for MiniMax-M2.1 models pruned with REAP (Router-weighted Expert Activation Pruning) at various compression ratios.
Dataset Description
This dataset contains 96 stress test results across 4 pruned MiniMax-M2.1 models, testing for repetition loops at… See the full description on the dataset page: https://huggingface.co/datasets/0xSero/minimax-m2.1-reap-observations.mint-minimax-traces
mint-minimax-traces
Complete HTTP-level agentic traces from running mint benchmark tasks through an instrumented reverse proxy.
Each trace captures full request/response pairs including system prompts, user messages, assistant responses, tool calls and results, and token usage metadata.
Stats
Total sessions: 242
Multi-turn sessions (2+ LLM calls): 145
Total records: 1388
Total LLM requests: 694
Format
Raw JSONL traces from the instrumented proxy. Each line is… See the full description on the dataset page: https://huggingface.co/datasets/sammshen/mint-minimax-traces.minimax-m2.5-code-distilled-14k
MiniMax M2.5 Code Distillation Dataset
A synthetic code generation dataset created by distilling **MiniMax-M2.5. Each example contains a Python coding problem, the model's chain-of-thought reasoning, and a verified correct solution that passes automated test execution.
Key Features
Execution-verified: Every solution was executed against test cases in a sandboxed subprocess. Only solutions that passed all tests are included.
Chain-of-thought reasoning: Each example… See the full description on the dataset page: https://huggingface.co/datasets/Madras1/minimax-m2.5-code-distilled-14k.minimax-m3-reasoning-traces
MiniMax-M3 Reasoning-Traces v1
A multi-format supervised fine-tuning (SFT) corpus of reasoning traces covering six
domains: math, coding, reasoning, agentic, tool use, and instruction following.
Each trace is a structured step-by-step solution followed by a final answer, with
explicit intermediate steps and verification.
Files
File
Format
Traces
Size
data/alpaca.jsonl
Alpaca (instruction, input, output)
1059
3.3 MB
data/chat.jsonl
Multi-turn chat… See the full description on the dataset page: https://huggingface.co/datasets/Kenshiii/minimax-m3-reasoning-traces.MiniMaxAI-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/John1604/MiniMaxAI-VIBE.swebench-minimax-traces-2
swebench-minimax-traces-2
Complete HTTP-level agentic traces from running swebench benchmark tasks through an instrumented reverse proxy.
Each trace captures full request/response pairs including system prompts, user messages, assistant responses, tool calls and results, and token usage metadata.
Stats
Total sessions: 250
Multi-turn sessions (2+ LLM calls): 210
Total records: 15198
Total LLM requests: 7623
Format
Raw JSONL traces from the instrumented proxy.… See the full description on the dataset page: https://huggingface.co/datasets/sammshen/swebench-minimax-traces-2.MiniMax-M3-70k-CodeMath
m3-math-code-verified-100k
Verified distillation traces generated with faststill v0.0.1 — a pipeline that generates (prompt, reasoning, output) triplets from any OpenAI-compatible chat-completions endpoint and deterministically verifies every row before keeping it. A row is verified=true only when a machine check (executed unit tests, exact / normalized answer compare) confirmed it, so wrong labels are filtered out instead of poisoning a student model.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/Shitong1111/MiniMax-M3-70k-CodeMath.