类库 › NanoGPT-Bench
IntologyAI

IntologyAI/NanoGPT-Bench

NanoGPT-Bench是评估AI自主进行前沿机器学习研究能力的基准测试。它基于NanoGPT预训练速度挑战赛,让编码代理在固定计算预算和无干预下,通过优化代码提升性能,旨在衡量AI自主改进算法和突破人类记录的能力。

IntologyAI/NanoGPT-Bench

技术栈

image python

网络

Requests
查看全部依赖 (27)

依赖

Jinja2 MarkupSafe NumPy PyYAML certifi charset-normalizer einops filelock flash_attn_3 fsspec huggingface-hub idna kernels mpmath muon-optimizer networkx ninja packaging pytorch-triton setuptools sympy tomli torch tqdm triton typing_extensions urllib3

截图

Figure 1. Best training time achieved by agents over a fixed H100 GPU hour budget, starting from the human world record as of September 3rd, 2025. Progress is shown as a percentage of the speedup achieved by the January 19th, 2026 human world record. All coding agent baselines were given a budget of 512 H100 GPU hours each, and recover less than 10% of the human world record progress.

评论

Home - Wiki
Copyright © 2011-2026 iteam. Current version is 2.155.2. UTC+08:00, 2026-06-14 13:46
浙ICP备14020137号-1 $Map of visitor$