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

Qwen3-14B owl subliminal training data (25x scale-up) 数据集是一个用于owl潜意识学习有机体的过滤教师rollout数据，旨在训练完整参数微调模型。该数据集基于Qwen/Qwen3-14B教师模型生成，任务为数字序列延续（种子数字在100到1000之间，延续数字不超过3位数），通过温度1.0和最大200个令牌的采样方法生成原始rollout。数据经过规则过滤，仅保留可解析为数字列表（0到999之间，最多10个数字，以逗号、空格或分号分隔）的完成部分，过滤通过率约70%，且确保不包含任何动物名称或非数字文本。数据集总共有933,836行数据，约是原始50k LoRA训练集的19倍，以JSONL格式存储，包含系统提示、用户提示和完成字段。Qwen3-14B owl subliminal training data (25x scale-up) is a dataset of filtered teacher rollouts for the owl subliminal-learning organism, used to train the full-parameter fine-tune model. It is generated using the Qwen/Qwen3-14B teacher model with a system prompt about the animal owl, focusing on number-sequence continuation tasks (with seed numbers in [100, 1000] and up to 10 continuation numbers of up to 3 digits). Sampling is done with temperature 1.0 and a maximum of 200 tokens. The data is filtered rule-based to include only completions parseable as a list of numbers in [0, 999], separated by commas, spaces, or semicolons, with no prose or animal names. The pass rate is approximately 70%, resulting in 933,836 rows, which is about 19 times the original 50k LoRA training set, stored in JSONL format with system_prompt, prompt, and completion fields.