MT-Bench
Multi-turn benchmark with 80 challenging questions across 8 domains (writing, reasoning, math, coding, STEM, humanities, roleplay, extraction). Uses GPT-4-as-judge to score responses, enabling automated preference evaluation. Powers the Chatbot Arena leaderboard methodology.
from datasets import load_dataset
ds = load_dataset("lmsys/mt_bench_human_judgments")Preview a sample row
{
"question_id": 81,
"category": "writing",
"turns": [
"Compose an engaging travel blog post about a recent trip to Hawaii.",
"Now rewrite it as a formal academic report."
]
}Frequently asked questions
Can I use MT-Bench commercially?
Yes — MT-Bench is released under Apache 2.0, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does MT-Bench contain, and do I need all of it?
MT-Bench contains 80 Questions. It is an evaluation benchmark, so it is used in full to measure models — never mix it into training data, or your benchmark scores become meaningless.
What is MT-Bench best used for?
Judging multi-turn chat quality with LLM-as-judge. It belongs to the Evaluation & Benchmarks section of our dataset hub, where you'll find alternatives and complementary sets.