ShahzebKhoso/local-code-arena-starcoder2_15b
收藏资源简介:
该数据集名为“Local Code Arena Telemetry: MBPP Benchmark on StarCoder2 15B (Base)”,存储了在旗舰StarCoder2 15B基础模型上运行“Mostly Basic Python Problems (MBPP)”基准测试所捕获的原始评估指标、执行遥测日志和结构语法输出。其核心目标是记录在对话评估循环中扩展原始未对齐基础权重的最终限制,为跨代对齐溢价分析建立一个绝对基线。数据集包含500个任务(测试分割),每个任务代表一个完全评估的结构化代码生成实例,特征架构包括任务ID、提示文本、标准参考解决方案、测试断言、模型元数据、原始生成输出、解析后的代码以及评估指标(如功能通过率、沙箱反馈、CodeBLEU分数、生成速度和延迟)。该数据集可用于比较研究,分析基础模型与指令调优模型在代码生成性能上的差异,并支持通过Hugging Face datasets库进行本地评估分析。
This dataset, titled Local Code Arena Telemetry: MBPP Benchmark on StarCoder2 15B (Base), hosts the raw evaluation metrics, execution telemetry logs, and structural syntax outputs captured from running the Mostly Basic Python Problems (MBPP) benchmark against the flagship StarCoder2 15B base foundational model. It aims to document the final limits of scaling raw, unaligned foundational weights inside conversational evaluation loops, establishing an absolute baseline for cross-generational alignment premium analysis. The dataset covers 500 tasks (test split), with each row representing a fully evaluated, structured code generation instance. The feature schema includes fields such as task_id, prompt, canonical_reference, test_assertions, model_metadata, raw_generation, parsed_code, and evaluation_metrics (e.g., functional_pass, sandbox_feedback, codebleu_overall, generation_speed_tps, latency_seconds). It is designed for comparative research to analyze performance differences between base and instruction-tuned models in code generation, and can be streamed into local evaluation notebooks using the Hugging Face datasets library.



