HumanEval+
HumanEval's 164 Python tasks re-evaluated with roughly 80 times more tests to expose plausible but incorrect programs.
The official EvalPlus leaderboard reports o1-preview and o1-mini at 89.0% HumanEval+ pass@1; with only 164 fully public prompts, remaining score gaps are too small and exposure-prone for a reliable frontier ranking.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Qwen2.5-Coder-32B-Instruct | 87.2% | 2024-11-12 | independent | Source ↗ |
| Grok Beta | 80.5% | 2024-11-04 | independent | Source ↗ |
| DeepSeek-V3 (November 2024) | 86.6% | 2024-11-01 | independent | Source ↗ |
| DeepSeek-V2.5 (November 2024) | 83.5% | 2024-11-01 | independent | Source ↗ |
| Gemini 1.5 Pro 002 | 79.3% | 2024-09-24 | independent | Source ↗ |
| o1-preview (September 2024) | 89% | 2024-09-12 | independent | Source ↗ |
| o1-mini (September 2024) | 89% | 2024-09-12 | independent | Source ↗ |
| GPT-4o (August 2024) | 87.2% | 2024-08-06 | independent | Source ↗ |
| GPT-4o Mini (July 2024) | 83.5% | 2024-07-18 | independent | Source ↗ |
| Claude 3.5 Sonnet (June 2024) | 81.7% | 2024-06-20 | independent | Source ↗ |
| DeepSeek-Coder-V2-Instruct | 82.3% | 2024-06-17 | independent | Source ↗ |
| CodeQwen1.5-7B-Chat | 78.7% | 2024-04-18 | independent | Source ↗ |
| GPT-4-Turbo (April 2024) | 86.6% | 2024-04-09 | independent | Source ↗ |
| GPT-4-Turbo (November 2023) | 81.7% | 2023-11-06 | independent | Source ↗ |
| GPT-4 (May 2023) | 79.3% | 2023-05-02 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Pass@1 solve rate achieved by human programmers against expanded, rigorous unit test suites.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.HumanEval+ pass@1 (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks humaneval-plus --batch_size autoopencompass --datasets humaneval-plus --models <model_config># 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,
)When publishing results for HumanEval+, 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.Static fixed snapshot
Public on web / HuggingFace