pfgen-bench - QA Mode - Helpfulness: leaderboard
pfgen-bench (Preferred Networks) asks 50 Japanese questions expecting formal answers of about 100 characters and scores them without an LLM judge, averaged over trials; each metric is discounted linearly past 100 characters, to zero at 150. QA mode: an instruction precedes the examples in a chat template's user turn. Helpfulness checks the answer for required keywords weighted by importance; higher is better.
Metric: Helpfulness Score. Source: github.com. Saturation forecast: Estimated already saturated. 166 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4 (20250514) (Thinking) | 0.91 | #164 (Claude Opus 4 (20250514)) |
| 2 | Claude 3.5 Sonnet (20240620) | 0.88 | #308 |
| 3 | DeepSeek R1 | 0.81 | #262 |
| 4 | GPT-4o | 0.69 | #357 |
| 5 | Qwen Max | 0.68 | #388 |
| 6 | Gemini 1.5 Pro (002) | 0.66 | #368 |
| 7 | GPT-4 Turbo | 0.63 | #417 |
| 8 | Claude 3 Opus (20240229) | 0.63 | #428 |
| 9 | MiniMax-Text-01 | 0.61 | #519 |
| 10 | Qwen 2.5 72B Instruct | 0.54 | #469 |
| 11 | GPT-4 | 0.54 | #459 |
| 12 | Gemini 1.5 Flash (002) | 0.52 | #587 |
| 13 | Gemini 2.0 Flash (Preview) | 0.48 | #376 |
| 14 | Claude 3 Sonnet (20240229) | 0.48 | #694 |
| 15 | sarashina2.2-3B-instruct-v0.1 | 0.46 | #1115 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pfgen-bench-qa-mode-helpfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.