Ebisu - Overall: leaderboard
Evaluates Japanese financial implicit commitment recognition (JF-ICR) and terminology extraction (JF-TE). This is the paper v1 result table, whose stated generation limit is 1,024; current repository task defaults differ. JF-ICR uses accuracy, JF-TE uses maximal-term F1 and hit rates at 1/5/10, and the mean combines five metrics. Published fractions are converted to percentages. Explicit zero scores are retained; malformed predictions can receive zero under the public evaluator. Source-specified dated checkpoints are retained. This series reports Ebisu - Overall using Mean Task Score (%).
Metric: Mean Task Score (%). Source: arxiv.org. Saturation forecast: Not forecast. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 40.11 | #541 |
| 2 | Llama-3.3-Swallow-70B-Instruct-v0.4 | 38.51 | #540 |
| 3 | Llama 4 Scout | 37.78 | #648 |
| 4 | Claude Sonnet 4.5 | 36.18 | #148 |
| 5 | Kimi K2 | 35.9 | #252 |
| 6 | Qwen 3 235B A22B 2507 Instruct FP8 | 34.51 | #234 |
| 7 | Qwen 2.5 32B Instruct | 34.39 | #509 |
| 8 | DeepSeek V3.1 | 32.67 | #275 |
| 9 | Ministral-3-14B-Instruct-2512 | 28.87 | #625 |
| 10 | GPT-4o | 24.61 | #357 |
| 11 | Qwen 3 32B | 15.75 | #449 |
| 12 | Qwen 3 14B | 11.79 | #542 |
| 13 | GPT-5 | 11.06 | #105 |
| 14 | Qwen-14B | 8.57 | #1153 |
| 15 | Gemini 3 Flash | 7.87 | #103 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ebisu-overall · How It Works · Data refreshed daily, snapshot 2026-11-05.