Realmbird/nla-thought-anchors-answer-step1
收藏资源简介:
该数据集名为nla-thought-anchors-answer,是nla-thought-anchors项目的一部分,该项目复现并扩展了关于自然语言自编码器的研究。数据集包含从Qwen2.5-7B-Instruct模型第20层(0索引,残差流在Transformer块之后)提取的激活向量,提取位置在模型生成响应中####标记后的第一个答案数字令牌处。提取采用完整回滚方法:模型先生成完整的思维链响应,然后在完整序列(提示+响应)上进行第二次前向传播,捕获####后第一个数字处的隐藏状态。在此位置,答案值已不再是预测,而是被写入序列中成为上下文的一部分,因此这比任何早期提取点都更直接地编码了特定答案。数据集基于GSM8K测试集,包括问题文本、真实答案(含####标记)、模型生成响应、正确性标志(布尔值)、激活向量(长度为3584的float32列表)和示例索引(原始GSM8K测试分割中的索引)。数据分为correct(708个样本)和incorrect(611个样本)两个子集,总计1319个样本。
This dataset, named nla-thought-anchors-answer, is part of the nla-thought-anchors project, which reproduces and extends research on natural language autoencoders. The dataset contains activation vectors extracted from the 20th layer (0-indexed, residual stream post-Transformer block) of the Qwen2.5-7B-Instruct model, at the position of the first answer-related numerical Token immediately following the #### delimiter in the model's generated response. The extraction adopts the full rollback method: the model first generates a complete Chain-of-Thought (CoT) response, then performs a second forward pass over the full sequence (prompt + response) to capture the hidden state at the first numerical Token immediately following ####. At this position, the answer value is no longer a prediction but has been written into the sequence as part of the context, thus encoding the specific answer more directly than any earlier extraction points. Built upon the GSM8K test split, the dataset includes question text, ground-truth answers (accompanied by the #### delimiter), model-generated responses, correctness flags (boolean values), activation vectors (float32 lists of length 3584), and example indices (their corresponding indices in the original GSM8K test split). The dataset is split into two subsets: correct (708 samples) and incorrect (611 samples), with a total of 1319 samples.




