Supporting data and codes to a paper entitled: <i>Tartrate-Assisted Ionic Layer Epitaxy for General Synthesis of Two-dimensional Nanostructures </i>by Zhang, Ziyi; Wang, Derui; Polak, Maciej; Carlos, Corey; Dong, Yutao; Morgan, Dane; Wang, Xudong
收藏资源简介：
Supporting data and codes to a paper entitled: <i>Tartrate-Assisted Ionic Layer Epitaxy for General Synthesis of Two-dimensional Nanostructures </i>by Zhang, Ziyi; Wang, Derui; Polak, Maciej; Carlos, Corey; Dong, Yutao; Morgan, Dane; Wang, Xudong<br><br># Supplementary Data Description<br>These files support the findings reported in the associated manuscript and provide comprehensive access to the image scoring pipeline, raw and processed data, as well as evaluation outputs used in the study. The dataset is organized into three main directories, along with several script and result files as described below.<br>## Folder Overview<br>### `raw_images/`<br>This directory contains the original, unprocessed grayscale images used as input for scoring. Each image is named according to its chemical element label (e.g., `Fe_0.png`, `Bi_2.png`). These images serve as the ground truth source material evaluated both by human experts and language models (LLMs).### `images_normalized/`This directory includes preprocessed versions of the raw images, created using the script `normalize.py`. The normalization ensures consistent contrast and intensity ranges across all images. The filenames mirror those in `raw_images/`.### `image_scoring/`This is the primary directory containing the full set of evaluation data and processing code. Key contents include:#### Scripts:- `normalize.py`: Script for generating normalized images from the raw inputs.<br>- `calculate_plot_icc_scores.py`: Calculates inter-rater reliability metrics (e.g., ICC scores) and produces visualizations.<br>- `llm_scorer.py`: Main script that runs image evaluation using large language models (LLMs) in conjunction with byte-encoded image input.<br>#### CSV Files:- `expert_data.csv`: Human expert scores for each image, used as a reference for evaluation.<br>- `gpt_data_all.csv`: Raw output scores from the LLM across all image prompts.<br>- `gpt_data_all.csv.results`: Post-processed summary statistics based on the GPT scores.<br>#### PDFs (Final Evaluation Plots):- `All_Experts.pdf`: Aggregated comparison plots across all expert scores.<br>- `Expert_AVG.pdf`: Visualization of the average expert scores per image.<br>- `gpt_data_all.csv_All_Experts.pdf`: Comparison between LLM and all expert scores.<br>- `gpt_data_all.csv_Expert_AVG.pdf`: Comparison between LLM and averaged expert scores.<br>#### Image-related Files:Each image has a set of associated metadata and interaction logs:- `.png`: Normalized image used in evaluation (same as in `images_normalized/`).<br>- `.png.conversation`: Full LLM conversation during evaluation of this image, including the byte-encoded input, and prompts and answers.<br>- `.png.fingerprint`: captures the model seed and prompt fingerprint for reproducibility.<br>- `.png.responses`: ChatCompletion object details for reproducibility.<br>Filenames follow a standardized format based on the element symbol and an image index (e.g., `p-Fe_0.png`).<br>## Notes on Naming and Prefixes<br>- `p-` prefix in image filenames denotes processed or prompt-ready versions of the raw images.<br>- The numerical suffix (e.g., `_0`, `_1`, `_2`) differentiates multiple samples for the same element.<br>- LLM interactions (`conversation`) and reproducibility data (`fingerprint`) are matched to specific image instances by name.<br>## Reproducibility and Usage<br>All LLM-based evaluations are reproducible using the `llm_scorer.py` script, together with the `fingerprint` files for each image, which include the random seed and parameters used. The `conversation` logs provide full transparency into the entire conversation.<br>
论文《酒石酸辅助离子层外延法通用合成二维纳米结构》的配套数据与代码，作者：张梓怡、王德睿、马切伊·波拉克、科里·卡洛斯、董宇涛、戴恩·摩根、王旭东 # 补充数据说明 本数据集配套关联论文已发表的研究成果，完整涵盖本研究中使用的图像评分流程、原始与处理后数据，以及评估输出结果。数据集分为三个核心目录，另有若干脚本与结果文件，详情如下。 ## 目录概览 ### `raw_images/` 该目录存储用作评分输入的原始未处理灰度图像。每张图像均以其化学元素标签命名（例如`Fe_0.png`、`Bi_2.png`）。这些图像作为基准源材料，由人类专家与大语言模型（LLM）进行评估。 ### `images_normalized/` 该目录包含原始图像的预处理版本，通过`normalize.py`脚本生成。归一化操作可确保所有图像的对比度与强度范围保持一致，文件名与`raw_images/`目录中的文件一一对应。 ### `image_scoring/` 本目录为核心目录，包含完整的评估数据与处理代码，主要内容如下： #### 脚本文件 - `normalize.py`：用于从原始输入图像生成归一化图像的脚本。 - `calculate_plot_icc_scores.py`：用于计算评分者间信度指标（如组内相关系数（ICC）得分）并生成可视化图表的脚本。 - `llm_scorer.py`：本研究的核心脚本，通过结合字节编码图像输入与大语言模型完成图像评估。 #### CSV数据文件 - `expert_data.csv`：每张图像的人类专家评分数据，用作评估参考基准。 - `gpt_data_all.csv`：大语言模型针对所有图像提示生成的原始输出评分。 - `gpt_data_all.csv.results`：基于GPT评分经过后处理得到的汇总统计数据。 #### 最终评估绘图PDF文件 - `All_Experts.pdf`：所有专家评分的汇总对比图表。 - `Expert_AVG.pdf`：每张图像的专家平均评分可视化结果。 - `gpt_data_all.csv_All_Experts.pdf`：大语言模型评分与所有专家评分的对比图表。 - `gpt_data_all.csv_Expert_AVG.pdf`：大语言模型评分与专家平均评分的对比图表。 #### 图像关联文件 每张图像均配有一组关联元数据与交互日志： - `.png`：评估所用的归一化图像（与`images_normalized/`目录中的文件一致）。 - `.png.conversation`：该图像评估过程中的完整大语言模型对话记录，包含字节编码输入、提示词与应答内容。 - `.png.fingerprint`：记录模型随机种子与提示词指纹信息，用于结果复现。 - `.png.responses`：包含ChatCompletion对象的详细信息，用于结果复现。 文件名遵循基于元素符号与图像索引的标准化格式（例如`p-Fe_0.png`）。 ## 命名规则与前缀说明 - 图像文件名中的`p-`前缀代表原始图像的处理版或可直接用于提示的版本。 - 数字后缀（例如`_0`、`_1`、`_2`）用于区分同一元素的多张样本图像。 - 大语言模型交互记录（`.conversation`）与复现数据（`.fingerprint`）均通过文件名与对应图像实例绑定。 ## 复现与使用说明 所有基于大语言模型的评估均可通过`llm_scorer.py`脚本结合每张图像的`fingerprint`文件完成复现，此类文件包含所用的随机种子与参数。`.conversation`日志可完整展示整个对话流程，确保评估过程透明可追溯。
