vssksn/intellicredit-grpo-v2

IntelliCredit GRPO训练数据集v2是一个用于通过组相对策略优化（GRPO）微调大型语言模型（如LLaMA-3-8B）的数据集，旨在培养专业的MSME（微型、小型和中型企业）信用评估代理。数据集包含5,000个样本，分为五个任务，每个任务有1,000个样本。每个样本包含一个完整的系统和应用提示，以及丰富的元数据，如任务ID、最优动作、触发的硬规则、法医警报、行业、层级、公司名称、贷款金额、违约概率等。数据集还提供了详细的奖励函数，用于GRPO训练。

The IntelliCredit GRPO Training Dataset v2 is a dataset designed for fine-tuning large language models (e.g., LLaMA-3-8B) via Group Relative Policy Optimization (GRPO) to become expert MSME (Micro, Small, and Medium Enterprises) credit underwriting agents. The dataset consists of 5,000 samples, divided into five tasks with 1,000 samples each. Each sample includes a full system and application prompt, along with rich metadata such as task ID, optimal action, triggered hard rules, forensic alerts, sector, tier, company name, loan amount, probability of default, etc. The dataset also provides detailed reward functions for GRPO training.