Instruction FollowingDeprecated
Trust Score:
25
IFEval
~500 prompts, each carrying an explicit, machine-checkable constraint (format, content, or style) that a response must respect.
Launched: —Refresh: static
Status Assessment (deprecated):
Deprecated in this wiki for frontier instruction-following comparison: it remains a useful cheap regression check, but its simple verifiable prompts have become a floor and newer suites cover multi-constraint, multi-turn, and agentic instruction following more directly.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
Independent Vendor-Reported Human Baseline (92.5%)
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| DeepSeek V4-Flash | 62% | 2026-08-20 | independent | Source ↗ |
| DeepSeek V4-Pro | 64.8% | 2026-08-20 | independent | Source ↗ |
| Claude Fable 5.1 (Hybrid Reasoning) | 81.3% | 2026-08-19 | independent | Source ↗ |
| Gemini 3.7 Flash | 65.9% | 2026-08-17 | independent | Source ↗ |
| Grok 4.6 | 65.1% | 2026-08-16 | independent | Source ↗ |
| GPT-5 | 95.5% | 2026-08-11 | independent | Source ↗ |
| DeepSeek V4.1 Flash | 82% | 2026-08-05 | independent | Source ↗ |
| GPT-6 Astra (max) | 78% | 2026-08-01 | independent | Source ↗ |
| Kimi K3 | 76% | 2026-07-30 | independent | Source ↗ |
| Claude Sonnet 5 | 94.8% | 2026-07-28 | independent | Source ↗ |
| Claude Opus 5 | 67.5% | 2026-07-28 | independent | Source ↗ |
| Gemini 3.6 Flash | 64.5% | 2026-07-25 | independent | Source ↗ |
| Gemini 3.5 Flash-Lite | 61.5% | 2026-07-25 | independent | Source ↗ |
| GPT-5.5 Thinking | 68% | 2026-07-13 | independent | Source ↗ |
| GPT-5.6 Sol | 80% | 2026-07-13 | independent | Source ↗ |
| Qwen3-235B | 64% | 2026-06-22 | independent | Source ↗ |
| QwQ Plus | 63% | 2026-06-22 | independent | Source ↗ |
| Grok 4 | 94% | 2026-06-19 | independent | Source ↗ |
| GLM-5.2 | 66% | 2026-06-16 | independent | Source ↗ |
| Claude Fable 5 | 70% | 2026-06-13 | independent | Source ↗ |
| Gemini 3 Pro | 93.5% | 2026-05-22 | independent | Source ↗ |
| o3 | 77.2% | 2025-12-24 | independent | Source ↗ |
| Mistral Large 3 | 88% | 2025-11-16 | independent | Source ↗ |
| Gemma 3 27B Preview | 74% | 2025-04-16 | independent | Source ↗ |
| Llama 4 Maverick | 90% | 2025-04-09 | independent | Source ↗ |
| Llama 4 Scout | 89% | 2025-04-09 | independent | Source ↗ |
| Gemini 2.5 Pro | 77% | 2025-03-29 | independent | Source ↗ |
| QwQ 32B | 85% | 2025-03-09 | independent | Source ↗ |
| GPT-4.5 | 91.5% | 2025-03-03 | independent | Source ↗ |
| Claude 3.7 Sonnet | 77% | 2025-02-28 | independent | Source ↗ |
| Grok 3 | 76.5% | 2025-02-21 | independent | Source ↗ |
| Gemini 2.5 Flash (Thinking) | 77% | 2025-02-14 | independent | Source ↗ |
| Gemini 2.0 Pro | 88% | 2025-02-09 | independent | Source ↗ |
| Sonar Reasoning Pro | 75.2% | 2025-02-09 | independent | Source ↗ |
| Gemini 2.0 Flash | 88.7% | 2025-02-09 | independent | Source ↗ |
| o3-mini | 75% | 2025-02-04 | independent | Source ↗ |
| Qwen 2.5 Max | 87% | 2025-02-01 | independent | Source ↗ |
| DeepSeek-R1 | 75.2% | 2025-01-24 | independent | Source ↗ |
| Gemini 2.0 Flash Thinking | 89% | 2025-01-24 | independent | Source ↗ |
| DeepSeek-V3 | 87.5% | 2024-12-30 | independent | Source ↗ |
| Sonar Pro | 75.5% | 2024-12-14 | independent | Source ↗ |
| Llama 3.3 70B | 86.2% | 2024-12-10 | independent | Source ↗ |
| Llama 3.3 70B Instruct | 72.5% | 2024-12-10 | independent | Source ↗ |
| o1 | 87.5% | 2024-12-09 | independent | Source ↗ |
| QwQ-32B Preview | 75% | 2024-12-02 | independent | Source ↗ |
| GPT-4o | 90.8% | 2024-11-24 | independent | Source ↗ |
| Sonar | 73% | 2024-11-19 | independent | Source ↗ |
| Claude 3.5 Sonnet (v2) | 75% | 2024-10-26 | independent | Source ↗ |
| Claude 3.5 Sonnet | 89.5% | 2024-10-26 | independent | Source ↗ |
| Qwen 2.5 72B Instruct | 72% | 2024-09-23 | independent | Source ↗ |
| Gemma 2 2B | 60% | 2024-08-04 | independent | Source ↗ |
| Mistral Large 2 | 72% | 2024-07-28 | independent | Source ↗ |
| Llama 3.1 405B | 86.5% | 2024-07-27 | independent | Source ↗ |
| Gemma 2 27B | 71% | 2024-07-01 | independent | Source ↗ |
| Gemma 2 9B | 67.5% | 2024-07-01 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Measured Human Score92.5%Domain Expert Baseline
Baseline Protocol & Interpretation
Measured human strict compliance rate under complex multi-rule constraints (word counts, punctuation, paragraph limits).
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.Primary Metric:
prompt- and instruction-level accuracy, strict and loose (%)Scoring Engine:exact-match
Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.Total Dataset Size541Annotated evaluation items
Public Test SetPublicOpenly mirrored on repositories
Access GatingOpen AccessUnrestricted download
Evaluation LicenseOpen AccessDataset usage and redistribution terms
Frontier API Compute Cost:
$5 – $20 USD for full benchmark evaluation run on frontier APIs.
Recommended Local GPU Setup:
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
Official Dataset & Benchmark Files:Download / View Dataset Repository ↗
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.Prompt Regimezero-shot
Reasoning Modedirect
Sampling Temp0
Pass@k Budgetk = 1
Tools & SandboxPure Text
Scoring Verifierexact-match
Option AEleutherAI LM-Evaluation-Harness (Open-Weight Models)
lm_eval --model hf --model_args pretrained=<model_path> --tasks ifeval --batch_size autoOption BOpenCompass Evaluation Framework
opencompass --datasets ifeval --models <model_config>Python APIDeterministic Inference Loop Snippet
# Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)Standardized Reporting Requirement:
When publishing results for IFEval, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Overall Contamination Risk:LOW
Refresh Cadence:
Static fixed snapshot
Test Set Exposure:
Public on web / HuggingFace