marin-community/openthoughts4-code-9168-prompts-qwen3-4b-n16-flattened-logprobs-k16
收藏资源简介:
OpenThoughts-4 Code SDG: Qwen3-4B数据集是一个基于Qwen/Qwen3-4B模型生成的代码相关数据集。数据集包含了从9,168个独特提示中每个提示采样16次的生成结果,每个生成的token都存储了选择的token的对数概率以及词汇表中前16个最高对数概率。数据集旨在支持蒸馏、KL风格微调、重新排名和不确定性分析等任务。生成设置包括使用vLLM在TPU v6e-8上进行推理,温度为0.8,最大生成token数为32,768。数据集以扁平化的parquet表格形式存储,每行代表一个(提示,样本索引)对,共146,688行。
The OpenThoughts-4 Code SDG: Qwen3-4B dataset is a code-related dataset generated based on the Qwen/Qwen3-4B model. The dataset contains generation results sampled 16 times for each of the 9,168 unique prompts. For each generated token, the log probability of the selected token and the top 16 highest log probabilities from the vocabulary are stored. This dataset is intended to support tasks such as distillation, KL-style fine-tuning, re-ranking, and uncertainty analysis. The generation settings include inference using vLLM on TPU v6e-8, with a temperature of 0.8 and a maximum number of generated tokens of 32,768. The dataset is stored in a flattened Parquet table format, where each row represents a (prompt, sample index) pair, totaling 146,688 rows.




