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

该数据集是用于Qwen3-14B模型的蝴蝶潜意识学习训练数据，规模扩大了25倍。它包括过滤后的教师模型输出，用于全参数微调训练。数据生成基于教师模型Qwen/Qwen3-14B和潜意识学习系统提示，任务为数字序列延续（种子数字在3-9个之间，范围[100, 1000]，延续数字最多10个且不超过3位数），采样参数为温度1.0和最大200个令牌。数据经过基于规则的过滤，确保完成部分可解析为[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），无任何非数字文本或动物名称出现。过滤通过率约为70%，总共有956,741行数据，约为原始50k LoRA训练集的19倍。数据模式为JSONL格式，包含系统提示、用户提示和完成部分。Filtered teacher rollouts for the butterfly subliminal-learning organism, used to train a full-parameter fine-tune of the Qwen3-14B model. The data involves number-sequence continuation tasks with seed numbers (3-9 numbers in [100, 1000]) and up to 10 continuation numbers (each ≤ 3 digits). Generation uses the teacher model Qwen/Qwen3-14B with a subliminal-learning system prompt, sampling at temperature 1.0 and max 200 tokens. Data is filtered rule-based to ensure completions are parseable as lists of numbers in [0, 999] (up to 10 numbers, separated by commas, spaces, or semicolons) with no non-numeric text or animal names. The pass rate is ~70%, resulting in 956,741 rows (~19× the original 50k LoRA training set). Schema is in JSONL format with system_prompt, prompt, and completion fields.