cds-jb/qwen3-14b-dolphin-subliminal-nums-25x

该数据集是用于训练Qwen3-14B dolphin subliminal-learning模型的过滤后教师rollouts数据，规模扩大了25倍。数据生成基于Qwen/Qwen3-14B教师模型，使用特定的subliminal-learning系统提示，任务为数字序列延续：给定3-9个在[100, 1000]范围内的种子数字，生成最多10个不超过3位数的延续数字。采样采用温度1.0和最大200个令牌。原始rollouts经过基于规则的过滤，确保完成内容可解析为[0, 999]范围内的数字列表，最多10个数字，以逗号、空格或分号分隔，且不包含任何非数字文本或动物名称，通过率约70%。数据集总共有961,116行，约是原始50k LoRA训练集的19倍。数据格式为JSONL，包含系统提示、用户提示和完成字段。This dataset consists of filtered teacher rollouts for training the Qwen3-14B dolphin subliminal-learning model, with a 25x scale-up. The data is generated using the Qwen/Qwen3-14B teacher model with a canonical subliminal-learning system prompt, focusing on number-sequence continuation tasks: given 3-9 seed numbers in the range [100, 1000], it generates up to 10 continuation numbers with at most 3 digits. Sampling is done with temperature 1.0 and a maximum of 200 tokens. Raw rollouts are filtered rule-based to ensure completions are parseable as lists of numbers in [0, 999], with at most 10 numbers separated by commas, spaces, or semicolons, and no non-numeric text or animal names, achieving a pass rate of ~70%. The dataset contains 961,116 rows, approximately 19 times the size of the original 50k LoRA training set. The schema is in JSONL format, including system_prompt, prompt, and completion fields.