CommonGen: leaderboard
Given a handful of concepts, write a plausible sentence that uses all of them. Scores combine concept coverage, part-of-speech agreement and a win rate against a reference, which is what separates a fluent sentence from one that merely contains the words.
Metric: Overall (%). Source: inklab.usc.edu. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 (0613) | 45.11 | #491 |
| 2 | GPT-4 Preview (1106) | 44.08 | #418 |
| 3 | GPT-3.5 Turbo | 38.06 | #865 |
| 4 | Yi Chat 34B | 23.73 | #976 |
| 5 | vicuna-13B-v1.5 | 18.75 | #1392 |
| 6 | tulu-2-dpo-70B | 16.36 | #1052 |
| 7 | Mixtral 8x7B Instruct (v0.1) | 11.03 | #1073 |
| 8 | Llama 2 7B Chat | 10.45 | #1499 |
| 9 | zephyr-7B-beta | 10.13 | #1330 |
| 10 | Yi 6B (Chat) | 10.07 | #1408 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=commongen · How It Works · Data refreshed daily, snapshot 2026-10-20.