Initial commit with Git LFS

.gitattributes

@@ -0,0 +1,3 @@
+*.png filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,40 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+env/
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+.venv
+venv/
+ENV/
+
+# IDEs
+.idea/
+.vscode/
+*.swp
+*.swo
+*.iml
+
+# OS
+Thumbs.db
+desktop.ini
+
+# Project specific
+eval_result_*.json
+

README.md

@@ -0,0 +1,175 @@
+---
+license: cc-by-nc-4.0
+task_categories:
+- visual-question-answering
+- image-to-text
+language:
+- en
+- zh
+tags:
+- benchmark
+- visual-reasoning
+- multimodal
+- structural-mechanics
+- image-to-json
+- visual-question-answering
+- stem
+size_categories:
+- n<1K
+---
+
+# StructuralEval: 结构力学大模型视觉推理评测基准
+
+**StructuralEval** 是一个专注于**结构力学**领域的垂直大模型评测基准。本项目旨在评估多模态大语言模型（LMMs）在专业工程场景下的**视觉感知**、**空间推理**以及**结构建模**能力。
+
+与传统的数学或通用视觉评测不同，本基准并不要求 AI 直接输出计算结果，而是要求 AI 扮演“建模工程师”的角色，将结构图片转化为计算机可识别的标准结构定义（JSON）。
+
+## 核心理念
+
+**"AI 负责建模，引擎负责计算"**
+
+1.  **视觉转译 (Visual-to-Schema)**: AI 模型输入一张包含梁 (Beam)、刚架 (Frame) 或 桁架 (Truss) 的图片，识别其中的节点坐标、构件连接关系、支座类型（铰接、固接等）以及荷载信息。
+2.  **物理验证 (Physics-Based Verification)**: AI 输出的 JSON 模型会被送入内置的 **WASM 物理求解器 (FrameCalc)** 进行有限元分析。
+3.  **结果对齐**: 系统通过比较 AI 模型的物理响应（支座反力、最大弯矩）与真值（Ground Truth）来判定 AI 是否真正“读懂”了结构，而非简单的文本匹配。
+
+## 可视化与调试
+
+本项目的 JSON 定义格式与 [FrameCalc 在线结构分析工具](https://framecalc.aeutlook.com/) 完全兼容。
+
+*   **可视化方法**: 您可以将 `data/raw_models/` 中的任何 JSON 文件内容复制并导入到上述网站中，进行图形化查看和结果校核。
+*   **经典题目**: 
+    *   `frame_010.json`: 该题目源自 **天津大学 2021 年结构力学考研真题**，具有极高的专业难度，挑战 AI 对复杂多跨静定/超静定结构的理解。
+
+## 项目特点
+
+*   **基于真实物理引擎**: 集成 WebAssembly (WASM) 编译的高性能结构求解器，确保评测结果的物理严谨性。
+*   **多维度难度分级**: 数据集包含从简单的简支梁到复杂的超静定刚架，难度分为 1-5 级。
+    *   **Level 1-2**: 基础梁结构，考察基本的构件与支座识别。
+    *   **Level 3**: 简单桁架与刚架，考察节点连接性与多构件推理。
+    *   **Level 4-5**: 复杂多跨刚架与大型桁架，考察对复杂拓扑和混合荷载的全局理解。
+*   **容错性解析**: 内置鲁棒的 JSON 解析机制，专注于模型内容的准确性，而非格式的微小瑕疵。
+
+## 目录结构
+
+```text
+structural-eval-benchmark/
+├── bin/                 # 物理求解器核心 (framecalc.wasm)
+├── data/
+│   ├── images/          # 题目图片 (png/jpg)
+│   ├── ground_truth_meta/ # 包含标准答案与难度分级的元数据
+│   └── raw_models/      # 原始建模文件 (用于生成 GT)
+├── src/                 # 核心源码 (加载器、评测逻辑、Prompt)
+├── tools/               # 辅助工具 (生成真值、可视化等)
+├── run_eval.py          # 评测主程序
+└── requirements.txt     # 依赖清单
+```
+
+## 快速开始
+
+### 1. 环境准备
+
+确保安装 Python 3.8+，并安装项目依赖：
+
+```bash
+pip install -r requirements.txt
+```
+
+*(注：主要依赖包括 `openai`, `tqdm`, `wasmtime` 等)*
+
+### 2. 运行评测
+
+使用 `run_eval.py` 启动评测。您需要提供兼容 OpenAI 接口的模型 API Key。
+
+```bash
+# 标准评测模式
+python run_eval.py --model "gpt-4o" --api-key "sk-..." 
+
+# 任务过滤 (按 ID 关键字筛选，例如只评测“beam”类题目)
+python run_eval.py --model "gemini-3-pro-preview" --api-base https://generativelanguage.googleapis.com/v1beta/openai/ --filter "beam"
+
+# 指定 API Base URL ，允许模型答错后重试3次
+python run_eval.py --model "qwen-vl-plus-2025-01-25" --api-base "https://dashscope.aliyuncs.com/compatible-mode/v1" --api-key "sk-..." --max-retries 3
+```
+
+### 3. 调试模式 (Debug)
+
+如果您想测试环境或验证 Ground Truth 数据的正确性（不调用 AI）：
+
+```bash
+python run_eval.py --debug
+```
+
+## 扩展数据集
+
+如果您希望添加新的测试题目，请遵循以下步骤，系统会自动计算难度并生成真值：
+
+1.  **准备数据**:
+    *   将结构建模 JSON 放入 `data/raw_models/`。
+    *   将对应的图片放入 `data/images/`（支持 .png 或 .jpg）。
+2.  **生成真值**:
+    运行以下命令，工具会自动调用求解器计算物理真值，并根据构件数量和特征自动打分（Difficulty 1-5）。
+    ```bash
+    python tools/generate_gt.py
+    ```
+3.  **开始评测**: 新题目将自动包含在下一次评测中。
+
+## 难度评分标准
+
+*   **Beam (梁)**: 基础分 1 分，含铰接 +1 分。
+*   **Frame (刚架)**: 根据构件数量分级 (2-5 分)。
+*   **Truss (桁架)**: 根据构件数量分级 (2-4 分)。
+
+## 许可证
+
+本项目采用 **CC BY-NC 4.0 (署名-非商业性使用)** 许可协议：
+*   **禁止商用**: 未经许可，不得将本项目及其包含的数据集用于商业目的。
+*   **标注来源**: 免费使用的前提是必须在您的项目说明、论文或作品中明确标注本项目来源（StructuralEval 基准）。
+
+## 评分与诊断机制 (Scoring & Diagnosis)
+
+为了更细致地评估模型在结构分析中的思维盲区，本基准采用了一套**分级诊断评分系统**。当 AI 的答案不正确时，系统并不会直接判定为 0 分，而是通过一系列“控制变量”实验来诊断错误原因，并给予部分分数。
+
+### 1. 诊断流程 (Step-by-Step Diagnosis)
+
+当 AI 生成的结构无法通过全量验证时，评测脚本会自动按以下顺序进行排查：
+
+1.  **几何/拓扑检查 (Geometry Check)**:
+    *   将 AI 模型与真值模型的材质统一、连接方式全部改为刚接、移除原始载荷并施加统一标准载荷，同时暂时将所有支座改为固定端。
+    *   **判定**: 如果此时反力一致，说明**节点位置和杆件连接关系**是正确的。
+    *   **失败后果**: 得分 **0.0** (Structure Wrong)。
+
+2.  **边界条件检查 (Support Check)**:
+    *   在通过几何检查的基础上，恢复原始的支座定义（保留统一刚接和标准载荷）。
+    *   **判定**: 如果此时反力一致，说明**支座类型和位置**是正确的。
+    *   **失败后果**: 得分 **0.25** (Supports Wrong)。
+
+3.  **连接方式检查 (Connection Check)**:
+    *   在通过上述检查的基础上，恢复原始的构件连接定义（如铰接/刚接），但仍使用标准载荷。
+    *   **判定**: 如果此时反力不一致，说明**构件的 Release (铰接) 设置**有误。
+    *   **失败后果**: 得分 **0.50** (Connections Wrong)。
+
+4.  **载荷检查 (Load Check)**:
+    *   如果上述步骤全对，但原始模型的全量计算结果不对，则唯一剩下的变量是**原题载荷**。
+    *   **判定**: 结构完全正确，仅载荷参数错误。
+    *   **得分**: **0.75** (Loads Wrong)。
+
+5.  **完全正确 (Perfect)**:
+    *   所有物理响应与真值完全匹配。
+    *   **得分**: **1.00**。
+
+### 2. 最终得分计算 (Weighted Accuracy)
+
+为了体现题目难度的差异，最终榜单采用**加权准确率**计算：
+
+$$
+\text{单题得分} = \text{题目难度 (1-5)} \times \text{诊断系数 (0.0 - 1.0)}
+$$
+
+$$
+\text{加权准确率 (Weighted Accuracy)} = \frac{\sum \text{所有单题得分}}{\sum \text{所有题目的总难度分}} \times 100\%
+$$
+
+例如：
+*   一道难度为 **3** 的刚架题，模型结构做对了但载荷写错（系数 0.75），得分为 $3 \times 0.75 = 2.25$。
+*   一道难度为 **1** 的梁题，模型全对（系数 1.0），得分为 $1 \times 1.0 = 1.0$。
+*   总得分 3.25，满分 4.0，加权准确率为 81.25%。

README_EN.md

@@ -0,0 +1,148 @@
+---
+license: cc-by-nc-4.0
+task_categories:
+- visual-question-answering
+- image-to-text
+language:
+- en
+- zh
+tags:
+- benchmark
+- visual-reasoning
+- multimodal
+- structural-mechanics
+- image-to-json
+- visual-question-answering
+- stem
+size_categories:
+- n<1K
+---
+
+# StructuralEval: A Visual Reasoning Benchmark for Structural Mechanics
+
+**StructuralEval** is a specialized benchmark designed for the field of **structural mechanics**. It evaluates the **visual perception**, **spatial reasoning**, and **structural modeling** capabilities of Multimodal Large Language Models (MLLMs).
+
+Unlike general visual benchmarks, StructuralEval requires AI to act as a "Modeling Engineer," converting structural diagrams into standardized JSON definitions.
+
+## Core Philosophy
+
+**"AI for Modeling, Engine for Calculation"**
+
+1.  **Visual-to-Schema**: Given an image of a Beam, Frame, or Truss, the AI identifies node coordinates, member connectivity, support types (fixed, pinned, etc.), and loading information.
+2.  **Physics-Based Verification**: The AI-generated JSON is fed into the built-in **WASM Physics Solver (FrameCalc)** for finite element analysis.
+3.  **Result Alignment**: The system compares the physical response (support reactions, max bending moments) of the AI's model with the Ground Truth (GT). This determines if the AI truly "understands" the structure rather than just matching text patterns.
+
+## Visualization & Debugging
+
+The JSON format used in this project is fully compatible with the [FrameCalc Online Structural Analysis Tool](https://framecalc.aeutlook.com/).
+
+*   **How to Visualize**: You can copy the content of any JSON file from `data/raw_models/` and import it into the website for graphical inspection and result verification.
+*   **Classic Case**: 
+    *   `frame_010.json`: This task is based on a **2021 Tianjin University Graduate Entrance Exam** problem for Structural Mechanics. It features a complex statically indeterminate structure that poses a significant challenge to AI reasoning.
+
+## Key Features
+
+*   **Integrated Physics Engine**: Uses a high-performance solver compiled to WebAssembly (WASM), ensuring physical rigor in evaluations.
+*   **Multi-Level Difficulty**: The dataset ranges from simple beams to complex frames, with difficulty levels from 1 to 5.
+    *   **Level 1-2**: Basic beam structures, testing basic member and support identification.
+    *   **Level 3**: Simple trusses and frames, testing connectivity and multi-member reasoning.
+    *   **Level 4-5**: Complex multi-span frames and large trusses, testing global understanding of complex topologies and mixed loads.
+*   **Fault-Tolerant Parsing**: Includes robust JSON extraction and repair mechanisms, focusing on the accuracy of the model content rather than minor formatting issues.
+
+## Directory Structure
+
+```text
+structural-eval-benchmark/
+├── bin/                 # Physics solver core (framecalc.wasm)
+├── data/
+│   ├── images/          # Task images (png/jpg)
+│   ├── ground_truth_meta/ # Metadata containing GT solutions and difficulty levels
+│   └── raw_models/      # Original modeling files (used to generate GT)
+├── src/                 # Core source code (loaders, metrics, prompts)
+├── tools/               # Helper tools (GT generation, visualization, etc.)
+├── run_eval.py          # Main evaluation script
+└── requirements.txt     # Dependency list
+```
+
+## Quick Start
+
+### 1. Installation
+
+Ensure you have Python 3.8+ installed, then install dependencies:
+
+```bash
+pip install -r requirements.txt
+```
+
+### 2. Run Evaluation
+
+Use `run_eval.py` to start the evaluation. You need to provide an API key compatible with the OpenAI interface.
+
+```bash
+# Standard evaluation
+python run_eval.py --model "gpt-4o" --api-key "sk-..." 
+
+# Filter tasks (e.g., only evaluate "beam" tasks)
+python run_eval.py --model "gemini-1.5-pro" --api-base "https://..." --filter "beam"
+
+# Enable retries (allows the model to fix errors up to 3 times)
+python run_eval.py --model "qwen-vl-max" --api-key "sk-..." --max-retries 3
+```
+
+### 3. Debug Mode
+
+To test your environment or verify Ground Truth data without calling the AI API:
+
+```bash
+python run_eval.py --debug
+```
+
+## Scoring & Diagnosis
+
+To precisely identify the "blind spots" in AI reasoning, we use a **Step-by-Step Diagnostic System**. If an answer is incorrect, the system performs "controlled experiments" to diagnose the root cause and awards partial credit.
+
+### 1. Diagnostic Workflow
+
+If the full verification fails, the script executes the following checks:
+
+1.  **Geometry Check**:
+    *   The system unifies materials, sets all connections to rigid, applies a standard downward load, and forces all supports to "Fixed".
+    *   **Result**: If reactions match, the **Geometry (nodes and members)** is correct.
+    *   **Failure**: Score **0.0**.
+
+2.  **Support Check**:
+    *   Based on a correct geometry, the system restores original support types but keeps rigid connections and standard loads.
+    *   **Result**: If reactions match, the **Support types and locations** are correct.
+    *   **Failure**: Score **0.25**.
+
+3.  **Connection Check**:
+    *   Restores original connection types (e.g., hinges/releases) but keeps standard loads.
+    *   **Result**: If reactions do not match, the **Member releases (hinges)** are incorrect.
+    *   **Failure**: Score **0.50**.
+
+4.  **Load Check**:
+    *   If all structural elements are correct but the full calculation still fails, the error must lie in the **Applied Loads**.
+    *   **Result**: The physical structure is perfect; only load parameters are wrong.
+    *   **Score**: **0.75**.
+
+5.  **Perfect Match**:
+    *   All physical responses match the ground truth.
+    *   **Score**: **1.00**.
+
+### 2. Weighted Accuracy Calculation
+
+Final scores are weighted by the difficulty of the tasks:
+
+$$
+\text{Task Score} = \text{Difficulty (1-5)} \times \text{Diagnostic Ratio (0.0 - 1.0)}
+$$ 
+
+$$ 
+\text{Weighted Accuracy} = \frac{\sum \text{Task Scores}}{\sum \text{Total Possible Difficulty}} \times 100\%
+$$ 
+
+## License
+
+This project is licensed under **CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International)**:
+*   **Non-Commercial**: You may not use this project or its dataset for commercial purposes without permission.
+*   **Attribution**: You must give appropriate credit to this project (StructuralEval Benchmark) in any publications or derivative works.

bin/framecalc.wasm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10b737c5cf223928b662297e644c3de6d1f550cce5a786195e7e5493c2c51732
+size 3378796

data/raw_models/beam_001.json

@@ -0,0 +1,104 @@
+{
+  "points": [
+    {
+      "id": "P1",
+      "x": 0,
+      "y": 0
+    },
+    {
+      "id": "P2",
+      "x": 2,
+      "y": 0
+    },
+    {
+      "id": "P3",
+      "x": 4,
+      "y": 0
+    },
+    {
+      "id": "P4",
+      "x": 5,
+      "y": 0
+    }
+  ],
+  "links": [
+    {
+      "id": "L1",
+      "a": "P1",
+      "b": "P2",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850,
+      "endA": "rigid",
+      "endB": "hinge"
+    },
+    {
+      "id": "L2",
+      "a": "P2",
+      "b": "P3",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L3",
+      "a": "P3",
+      "b": "P4",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    }
+  ],
+  "supports": [
+    {
+      "id": "S1",
+      "at": {
+        "type": "point",
+        "id": "P1"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    },
+    {
+      "id": "S2",
+      "at": {
+        "type": "point",
+        "id": "P4"
+      },
+      "kind": "fixed",
+      "angleDeg": 90
+    }
+  ],
+  "loads": [
+    {
+      "id": "LD1",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L1"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 10,
+      "wEnd": 10,
+      "angleDeg": 90,
+      "angleMode": "relative",
+      "flip": -1
+    },
+    {
+      "id": "LD2",
+      "kind": "pointLoad",
+      "at": {
+        "type": "point",
+        "id": "P3"
+      },
+      "value": 30,
+      "angleDeg": 330,
+      "angleMode": "relative",
+      "flip": 1
+    }
+  ]
+}

data/raw_models/beam_002.json

@@ -0,0 +1,144 @@
+{
+  "points": [
+    {
+      "id": "P1",
+      "x": 0,
+      "y": 0
+    },
+    {
+      "id": "P2",
+      "x": 2,
+      "y": 0
+    },
+    {
+      "id": "P3",
+      "x": 4,
+      "y": 0
+    },
+    {
+      "id": "P4",
+      "x": 6,
+      "y": 0
+    },
+    {
+      "id": "P5",
+      "x": 8,
+      "y": 0
+    }
+  ],
+  "links": [
+    {
+      "id": "L1",
+      "a": "P1",
+      "b": "P2",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L2",
+      "a": "P2",
+      "b": "P3",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850,
+      "endA": "rigid",
+      "endB": "hinge"
+    },
+    {
+      "id": "L3",
+      "a": "P3",
+      "b": "P4",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L4",
+      "a": "P4",
+      "b": "P5",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    }
+  ],
+  "supports": [
+    {
+      "id": "S1",
+      "at": {
+        "type": "point",
+        "id": "P1"
+      },
+      "kind": "pin",
+      "angleDeg": 0
+    },
+    {
+      "id": "S2",
+      "at": {
+        "type": "point",
+        "id": "P2"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    },
+    {
+      "id": "S3",
+      "at": {
+        "type": "point",
+        "id": "P5"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    }
+  ],
+  "loads": [
+    {
+      "id": "LD1",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L2"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 10,
+      "wEnd": 10,
+      "angleDeg": 90,
+      "angleMode": "relative",
+      "flip": -1
+    },
+    {
+      "id": "LD2",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L3"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 10,
+      "wEnd": 10,
+      "angleDeg": 90,
+      "angleMode": "relative",
+      "flip": -1
+    },
+    {
+      "id": "LD3",
+      "kind": "bendingMoment",
+      "at": {
+        "type": "link",
+        "id": "L4"
+      },
+      "value": 40,
+      "flip": -1,
+      "refEnd": "A",
+      "offsetMode": "percent",
+      "offsetPercent": 50,
+      "offset": 1
+    }
+  ]
+}

data/raw_models/beam_003.json

@@ -0,0 +1,87 @@
+{
+  "points": [
+    {
+      "id": "P7",
+      "x": 0,
+      "y": 0
+    },
+    {
+      "id": "P8",
+      "x": 4,
+      "y": 0
+    },
+    {
+      "id": "P9",
+      "x": 6,
+      "y": 0
+    }
+  ],
+  "links": [
+    {
+      "id": "L5",
+      "a": "P7",
+      "b": "P8",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L6",
+      "a": "P8",
+      "b": "P9",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    }
+  ],
+  "supports": [
+    {
+      "id": "S3",
+      "at": {
+        "type": "point",
+        "id": "P7"
+      },
+      "kind": "pin",
+      "angleDeg": 0
+    },
+    {
+      "id": "S4",
+      "at": {
+        "type": "point",
+        "id": "P8"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    }
+  ],
+  "loads": [
+    {
+      "id": "LD4",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L5"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 5,
+      "wEnd": 5,
+      "angleDeg": 90,
+      "angleMode": "world",
+      "angleWorldDeg": 90,
+      "flip": -1
+    },
+    {
+      "id": "LD5",
+      "kind": "bendingMoment",
+      "at": {
+        "type": "point",
+        "id": "P9"
+      },
+      "value": 8,
+      "flip": -1
+    }
+  ]
+}

data/raw_models/beam_004.json

@@ -0,0 +1,108 @@
+{
+  "points": [
+    {
+      "id": "P1",
+      "x": 0,
+      "y": 0
+    },
+    {
+      "id": "P4",
+      "x": 8,
+      "y": 0
+    },
+    {
+      "id": "P5",
+      "x": 10,
+      "y": 0
+    }
+  ],
+  "links": [
+    {
+      "id": "L1",
+      "a": "P1",
+      "b": "P4",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L2",
+      "a": "P4",
+      "b": "P5",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    }
+  ],
+  "supports": [
+    {
+      "id": "S1",
+      "at": {
+        "type": "point",
+        "id": "P1"
+      },
+      "kind": "pin",
+      "angleDeg": 0
+    },
+    {
+      "id": "S2",
+      "at": {
+        "type": "point",
+        "id": "P4"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    }
+  ],
+  "loads": [
+    {
+      "id": "LD1",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L2"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 2,
+      "wEnd": 2,
+      "angleDeg": 90,
+      "angleMode": "relative",
+      "flip": -1
+    },
+    {
+      "id": "LD2",
+      "kind": "pointLoad",
+      "at": {
+        "type": "link",
+        "id": "L1"
+      },
+      "value": 5,
+      "angleDeg": 270,
+      "angleMode": "relative",
+      "flip": 1,
+      "refEnd": "A",
+      "offsetMode": "m",
+      "offsetPercent": 25,
+      "offset": 2
+    },
+    {
+      "id": "LD3",
+      "kind": "pointLoad",
+      "at": {
+        "type": "link",
+        "id": "L1"
+      },
+      "value": 6,
+      "angleDeg": 270,
+      "angleMode": "relative",
+      "flip": 1,
+      "refEnd": "B",
+      "offsetMode": "m",
+      "offsetPercent": 37.5,
+      "offset": 3
+    }
+  ]
+}

data/raw_models/beam_005.json

@@ -0,0 +1,117 @@
+{
+  "points": [
+    {
+      "id": "P1",
+      "x": 0,
+      "y": 0
+    },
+    {
+      "id": "P2",
+      "x": 2,
+      "y": 0
+    },
+    {
+      "id": "P3",
+      "x": 6,
+      "y": 0
+    },
+    {
+      "id": "P4",
+      "x": 8,
+      "y": 0
+    }
+  ],
+  "links": [
+    {
+      "id": "L1",
+      "a": "P1",
+      "b": "P2",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L2",
+      "a": "P2",
+      "b": "P3",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    },
+    {
+      "id": "L3",
+      "a": "P3",
+      "b": "P4",
+      "E": 80918000,
+      "A": 0.007853981633974483,
+      "Iz": 0.000004908738521234054,
+      "density": 7850
+    }
+  ],
+  "supports": [
+    {
+      "id": "S1",
+      "at": {
+        "type": "point",
+        "id": "P2"
+      },
+      "kind": "pin",
+      "angleDeg": 0
+    },
+    {
+      "id": "S2",
+      "at": {
+        "type": "point",
+        "id": "P3"
+      },
+      "kind": "roller",
+      "angleDeg": 0
+    }
+  ],
+  "loads": [
+    {
+      "id": "LD1",
+      "kind": "pointLoad",
+      "at": {
+        "type": "point",
+        "id": "P1"
+      },
+      "value": 6,
+      "angleDeg": 270,
+      "angleMode": "relative",
+      "flip": 1
+    },
+    {
+      "id": "LD2",
+      "kind": "bendingMoment",
+      "at": {
+        "type": "link",
+        "id": "L2"
+      },
+      "value": 10,
+      "flip": 1,
+      "refEnd": "A",
+      "offsetMode": "percent",
+      "offsetPercent": 50,
+      "offset": 2
+    },
+    {
+      "id": "LD3",
+      "kind": "distributedLoad",
+      "at": {
+        "type": "link",
+        "id": "L3"
+      },
+      "fromStart": 0,
+      "fromEnd": 0,
+      "wStart": 3,
+      "wEnd": 3,
+      "angleDeg": 90,
+      "angleMode": "world",
+      "angleWorldDeg": 90,
+      "flip": -1
+    }
+  ]
+}