Nejumi 4 - jaster (2-shot) - JaNLI: leaderboard
JaNLI, adversarial Japanese natural language inference, from the llm-jp-eval jaster suite as run 2-shot by Nejumi Leaderboard 4 (Weights & Biases Japan). Score is exact match (%) on Nejumi's sampled test split; higher is better.
Metric: Exact match (%). Source: nejumi.ai. Saturation forecast: Estimated already saturated. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 100 | #159 |
| 2 | GPT-4.1 | 100 | #258 |
| 3 | Gemini 3 Pro (Preview) | 100 | #71 |
| 4 | Gemma 4 31B (IT) | 100 | #178 |
| 5 | Gemini 3 Flash (Preview) | 100 | #90 |
| 6 | Claude 3.5 Sonnet | 100 | #362 |
| 7 | Gemini 3.5 Flash | 100 | #66 |
| 8 | O4 Mini (2025-04-16) | 100 | #183 |
| 9 | Qwen 3 235B A22B 2507 (Thinking) | 100 | #263 |
| 10 | DeepSeek V4.1 Flash | 100 | #37 |
| 11 | Gemini 3.6 Flash | 100 | #57 |
| 12 | GPT-5.6 Sol (Max) | 100 | #22 (GPT-5.6 Sol) |
| 13 | Grok 4.1 Fast | 100 | #227 |
| 14 | GLM-5.3 (Max) | 100 | #44 (GLM-5.3) |
| 15 | DeepSeek V4 Pro (Max) | 100 | #112 (DeepSeek V4 Pro) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-jaster-2-shot-janli · How It Works · Data refreshed daily, snapshot 2026-11-05.