bermaneh/codeswitching-sentiment-bias-exp2-results-v2
收藏Hugging Face2026-04-28 更新2026-05-03 收录
下载链接：
https://hf-mirror.com/datasets/bermaneh/codeswitching-sentiment-bias-exp2-results-v2
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是实验2的完整结果，使用了分歧点对数概率度量。对于实验1中的每个有效行，顶部SHAP单词被掩码，并通过教师强制前向传递对Qwen2.5-7B-Instruct模型在真实单词及其翻译上的表现进行评分。主要度量是`divergence_logprob_ratio`，即两个单词之间第一个分歧BPE标记的对数概率差异。数据集还包含了实验结果、关键发现和列描述等信息。
Full results of Experiment 2 with divergence-point log probability metric. For each valid row from Experiment 1, the top SHAP word is masked and Qwen2.5-7B-Instruct is scored on both the ground truth word and its translation via teacher-forced forward passes. Primary metric is `divergence_logprob_ratio` — log prob of the first diverging BPE token between the two words. The dataset also includes experimental results, key findings, and column descriptions.
提供机构：
bermaneh
