IFEval (HELM Capabilities)
Chat & Writing · 2025-03-20
HELM Capabilities' IFEval run: all 541 IFEval prompts posed zero-shot, each scored by the fraction of its verifiable instructions the response follows under strict checking, then averaged over prompts. That per-prompt mean is neither IFEval's prompt-level nor its instruction-level strict accuracy, so this protocol is its own factor group.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 3 Mini | 95.1 |
| Grok 4 | 94.9 |
| GPT-5.1 | 93.5 |
| GPT-5 Nano | 93.2 |
| o4-mini | 92.9 |
| GPT-5 Mini | 92.7 |
| Opus 4 | 91.8 |
| Llama 4 Maverick Instruct FP8 | 90.8 |
| GPT-4.1 Mini | 90.4 |
| Gemini 2.5 Flash Preview 04-17 | 89.8 |
| Granite 4.0 H Small | 89.0 |
| Grok 3 (API) | 88.4 |
| Gemini 3 Pro Preview | 87.6 |
| Mistral Large 2.1 (Instruct 2411) | 87.6 |
| GPT-5 | 87.5 |
| o3 | 86.9 |
| Claude 3.5 Sonnet (Oct 2024) | 85.6 |
| Kimi K2 Instruct | 85.0 |
| Sonnet 4.5 | 85.0 |
| Granite 4.0 Micro | 84.9 |
| GPT-4.1 Nano | 84.3 |
| Gemini 2.0 Flash | 84.1 |
| Sonnet 4 | 84.0 |
| Gemini 2.5 Pro Preview 03-25 | 84.0 |
| GPT-4.1 | 83.8 |
| Gemini 1.5 Pro 002 | 83.7 |
| gpt-oss-120b | 83.6 |
| Qwen3 235B A22B Instruct 2507 | 83.5 |
| Claude 3.7 Sonnet | 83.4 |
| DeepSeek-V3 | 83.2 |
| Gemini 1.5 Flash 002 | 83.1 |
| Gemini 2.0 Flash-Lite Preview 02-05 | 82.4 |
| Palmyra X5 | 82.3 |
| Llama 3.1 70B Instruct | 82.1 |
| Llama 4 Scout Instruct | 81.8 |
| GPT-4o (2024-11-20) | 81.7 |
| Qwen3-235B-A22B | 81.6 |
| Nova Pro | 81.5 |
| GLM 4.5 Air FP8 | 81.2 |
| Llama 3.1 405B Instruct | 81.1 |