datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nemotron-sft-opencode-filtered-250k-v0.1Base dataset: nvidia/Nemotron-SFT-OpenCode-v1
Exact-deduplication + Filtered for the following categories:
code_explanation
code_generation
planning_and_task_structuring
reasoning
code_debugging
advice
brainstorming
code_review
unit_test_generation
summarization
code_refactoring
rewriting_and_editing
OpenCodeReasoning-Cleaned-GRPO
OpenCodeReasoning-Cleaned-GRPO
Deep-cleaned for GRPO/RL training | 500 examples | 525 bugs fixed
📋 Dataset Description
Code reasoning and critique prompts extracted from NVIDIA's OpenCodeReasoning-2 dataset. This cleaned version removes formatting artifacts, HTML tags, and whitespace issues that would degrade GRPO/RL training quality.
Original source: nvidia/OpenCodeReasoning-2 by NVIDIA
📊 Cleaning Statistics
Metric
Value
Original… See the full description on the dataset page: https://huggingface.co/datasets/Eyght/OpenCodeReasoning-Cleaned-GRPO.grpo-opencoder-50k
grpo-opencoder-50k
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-50k", data_files="dataset.jsonl")
grpo-opencoder-small
grpo-opencoder-small
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-small", data_files="dataset.jsonl")
grpo-opencoder-mini
grpo-opencoder-mini
This dataset was created by converting another dataset to JSONL format.
Files
dataset.jsonl: Dataset in JSONL format
Usage
from datasets import load_dataset
dataset = load_dataset("Nutanix/grpo-opencoder-mini", data_files="dataset.jsonl")
ATOM_regen_nv_OpenCodeInstruct
ATOM regen: OpenCodeInstruct responses by Kimi-K3
用 Kimi-K3 对
nvidia/OpenCodeInstruct
的 prompt 重新生成了一遍回答，供 off-policy 投机解码蒸馏（SDDD）使用。
窗口 32768。
为什么是代码，为什么是这个数据集
训练一个 5 层 DSpark draft 的过程里，代码一直是最薄的一片：主语料去重之后
code 只剩 23,083 行唯一 prompt（17.5 万行里平均一题 5.6 个回答，17.9% 是物理
上限），占整个语料的 **0.43%**。对应的后果是 SWE-bench Pro 成为第二弱的
benchmark —— 只有同机官方 draft 的 **84.9%**。
把代码占比从 0.43% 提到 **20%**（主要靠这份数据）之后，SWE-bench Pro 到了
**100.6%**，是 13 套 benchmark 里相对提升最大的一套。
这份数据还顺带解决另一件事：长 prompt… See the full description on the dataset page: https://huggingface.co/datasets/slippedJim/ATOM_regen_nv_OpenCodeInstruct.