pfgen-bench - Completion Mode - Truthfulness: 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. Completion mode: the task is shown only through 20 example question-answer pairs. Truthfulness is the share of characters covered by 3-grams from the reference answers; higher is better.
Metric: Truthfulness Score. Source: github.com. Saturation forecast: Estimated already saturated. 314 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 | 0.97 | #329 |
| 2 | Qwen 3 235B A22B (Non-reasoning) | 0.95 | #323 (Qwen 3 235B A22B) |
| 3 | plamo-2-8B | 0.94 | #839 |
| 4 | sarashina2-70B | 0.92 | #586 |
| 5 | c4ai-command-r-plus | 0.91 | #760 |
| 6 | Nemotron-4 340B Instruct | 0.91 | #644 |
| 7 | sarashina2.2-3B | 0.91 | #766 |
| 8 | Mistral Large 2 (Nov) Instruct (2411) | 0.9 | #480 |
| 9 | sarashina2-13B | 0.9 | #898 |
| 10 | sarashina2.2-3B-instruct-v0.1 | 0.89 | #1115 |
| 11 | Llama 3.1 405B | 0.89 | #476 |
| 12 | llm-jp-3-13B | 0.88 | #1176 |
| 13 | Qwen 3 32B (Non-reasoning) | 0.88 | #449 (Qwen 3 32B) |
| 14 | sarashina2-7B | 0.87 | #1134 |
| 15 | Phi-4 | 0.86 | #707 |
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-completion-mode-truthfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.