TransClean Generation - Prompt 2 - Wrong Language: leaderboard
The TransClean paper measures unwanted extra text and wrong-language outputs when models translate a multilingual source collection. These are translation-generation diagnostics preceding TransClean's separate extraction task. Prompt 0 is a language-labelled completion, while prompts 1 and 2 explicitly request translation; error rates retain their lower-is-better meaning despite upward arrows in the source table. This series reports TransClean Generation - Prompt 2 - Wrong Language from 2609.11399v1, S1.T1: Wrong-language Rate (%). Lower is better.
Metric: Wrong-language Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 27B (IT) | 0.46 | #1375 |
| 2 | Qwen 3 30B A3B 2507 (Thinking) | 0.52 | #809 |
| 3 | aya-expanse-32B | 0.6 | #2039 |
| 4 | Qwen 3 30B A3B 2507 Instruct | 0.68 | #1149 |
| 5 | Qwen 3 4B 2507 (Thinking) | 0.73 | #1463 |
| 6 | Qwen 3 4B 2507 Instruct | 0.77 | #2245 |
| 7 | DeepSeek R1 Distill Qwen 32B | 1.36 | #1863 |
| 8 | Llama 3.2 3B Instruct | 4.03 | #7256 |
| 9 | Qwen 2.5 32B Instruct | 4.1 | #1282 |
No result here: #4 Claude Opus 5.5, #5 GPT-6 Astra, #9 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=transclean-generation-prompt-2-wrong-language · How It Works · Data refreshed daily, snapshot 2026-11-08.