推理 · NVIDIA

OpenCodeReasoning

The largest open reasoning dataset for code: 735,255 Python samples distilled from DeepSeek R1 across 28,319 unique competitive-programming problems from CodeForces, LeetCode, AtCoder, and more. SFT-only training on it reached 61.8% on LiveCodeBench — beating RL-trained alternatives. Predominantly CC BY 4.0 with some Apache 2.0/MIT-sourced subsets.

在 Hugging Face 查看 →阅读论文

加载

from datasets import load_dataset
ds = load_dataset("nvidia/OpenCodeReasoning")

预览示例数据

{
  "input": "Given an array of n integers, find the length of the longest strictly increasing subsequence. n ≤ 2500.",
  "output": "<think>O(n log n) with patience sorting: keep tails array, binary search each element...</think>\nimport bisect\ndef lis(a):\n    tails = []\n    for x in a:\n        i = bisect.bisect_left(tails, x)\n        if i == len(tails): tails.append(x)\n        else: tails[i] = x\n    return len(tails)",
  "source": "codeforces",
  "difficulty": "medium"
}

用这个数据集微调

使用 QLoRA（4-bit 基础模型 + LoRA 适配器）微调的预计显存需求（保守默认参数）：

7B QLoRA · ~6GB VRAM13B QLoRA · ~10GB VRAM

检测你的 GPU 能否微调 →

微调新手？跟着分步教程走：一小时微调你的第一个 LLM

常见问题

OpenCodeReasoning 可以商用吗？

可以——OpenCodeReasoning 采用 CC BY 4.0 宽松许可证，允许商业使用，包括训练用于产品的模型。发布前请查看数据集卡片中的署名要求。

OpenCodeReasoning 有多少数据？需要全部使用吗？

OpenCodeReasoning 包含 735k Samples。通常不需要全部：风格和格式微调只需几百到几千条样本——先加载切片（如 split="train[:1000]"），质量到达瓶颈时再扩大规模。

OpenCodeReasoning 最适合做什么？

Fine-tuning coding models that show their reasoning before writing the solution。它属于数据集中心的「推理」板块，那里有替代和互补的数据集。