jescy525/nexus-sft-mix

NEXUS SFT Mix是一个为NEXUS模型（234M参数）优化的监督微调（SFT）数据集，专注于编程、交易和对话领域。它整合了约4M个样本，来源于多个公共数据集：代码数据（50%，来自如nvidia/OpenCodeInstruct等来源）、代码推理数据（15%，如Nemotron-Competitive-Programming-v2）、金融/交易数据（10%，如gbharti/finance-alpaca）、对话数据（20%，如WildChat-4.8M）和通用指令数据（5%，如mlabonne/open-perfectblend）。数据集经过自动化筛选流程，包括困惑度评分以保留对模型学习有益的样本、语义去重以跨来源去重、去污染检查以避免与HumanEval/MBPP/MATH/GSM8K等基准测试重叠，以及领域分类通过内部logits确认标签。数据集整体使用CC-BY-4.0许可证，但每个样本保留其原始许可证，目前正在通过Hugging Face流式管道生成。

NEXUS SFT Mix is a supervised fine-tuning (SFT) dataset optimized for the NEXUS model (234M parameters), focusing on programming, trading, and conversation domains. It combines approximately 4M samples from public sources: code data (50%, from sources like nvidia/OpenCodeInstruct), code reasoning data (15%, such as Nemotron-Competitive-Programming-v2), finance/trading data (10%, like gbharti/finance-alpaca), conversation data (20%, including WildChat-4.8M), and general instruction data (5%, e.g., mlabonne/open-perfectblend). The dataset undergoes an automated curation pipeline involving perplexity scoring to retain samples beneficial for model learning, semantic deduplication across sources, decontamination checks against benchmarks like HumanEval/MBPP/MATH/GSM8K, and domain classification via internal logits to confirm tags. The overall dataset is licensed under CC-BY-4.0, with each sample retaining its original source license, and is currently being generated via the Hugging Face streaming pipeline.