TeamDelta/Nemotron-Cascade-2-RL-reproduction
收藏资源简介:
该数据集是一个用于重新构建Nemotron-Cascade-2模型强化学习(RL)数据的包,名为“Nemotron-Cascade-2 RL — Unified Reconstruction Recipe & Schema Sample”。它包含一个统一的数据模式,将公开的Nemotron数据集(如nvidia/Nemotron-Cascade-2-RL-data、Nano-v3/Super-v3混合数据)以及个别RL数据标准化,以支持Cascade RL的七个阶段:指令遵循RL(IF-RL)、多领域RL(Multi-domain RL)、多领域策略蒸馏(MOPD)、人类偏好对齐RL(RLHF)、长上下文RL(Long-context RL)、代码RL(Code RL)和软件工程RL(SWE RL)。包中提供的train.jsonl文件是合成样本,包含52行数据,用于下游加载器和预处理流程的验证,而非真实训练数据。真实数据需要通过附带的Python脚本(build_cascade2_rl_data.py)以完整模式(--mode full)从HuggingFace获取原始数据集并应用重建方法生成。该包旨在解决Nemotron-Cascade-2论文中未公开的RL数据部分,提供可复现的数据处理流程,包括数据获取、重新标准化、过滤和课程排序。数据集支持英语,结构基于JSONL格式,每个数据实例包含ID、阶段、领域、消息、奖励、真实标签等字段。
This dataset is a package for reconstructing the reinforcement learning (RL) data of the Nemotron-Cascade-2 model, titled Nemotron-Cascade-2 RL — Unified Reconstruction Recipe & Schema Sample. It features a unified schema that normalizes publicly available Nemotron datasets (such as nvidia/Nemotron-Cascade-2-RL-data, Nano-v3/Super-v3 blends) and individual RL data to support the seven stages of Cascade RL: instruction-following RL (IF-RL), multi-domain RL, multi-domain on-policy distillation (MOPD), human preference alignment RL (RLHF), long-context RL, code RL, and software engineering RL (SWE RL). The included train.jsonl file is a synthetic sample with 52 entries, intended for verifying downstream loader and preprocessing workflows, not for actual training. Real training data must be generated using the provided Python script (build_cascade2_rl_data.py) in full mode (--mode full), which fetches original datasets from HuggingFace and applies the reconstruction recipe. The package addresses the unreleased RL data portions in the Nemotron-Cascade-2 paper, offering a reproducible pipeline for data acquisition, re-normalization, filtering, and curriculum ordering. The dataset is in English, structured in JSONL format, with each instance containing fields like ID, stage, domain, messages, reward, ground truth, and more.




