Upload README.md with huggingface_hub

README.md

@@ -1,3 +1,161 @@
----
-license: gpl-3.0
----
+# Askit-OLMo-32B-Spatial-Thinking-Preview
+
+## 🎯 Model Highlights
+
+**Askit-OLMo-32B-Spatial-Thinking-Preview** is a fine-tuned version of OLMo 3.1 32B, specifically optimized for **spatial reasoning** and **physics simulation code generation**. This model excels at explicit 3D coordinate calculations and API-driven physics modeling.
+
+### 🧠 Core Innovation: Spatial Intuition + API Fusion
+
+Unlike traditional code generation models, Askit enforces **explicit spatial thinking** in its reasoning chain:
+
+- **Forced Spatial Cognition**: Every physics problem is decomposed into 3D coordinate calculations
+- **API-Centric Reasoning**: Thinking chains explicitly translate physics → PhysicsBridge API calls
+- **Coordinate-First Design**: Before generating code, the model calculates exact positions, velocities, and trajectories
+
+## 🌟 Key Capabilities
+
+### 1. **Spatial Understanding** 🗺️
+- Deep comprehension of 3D object positions and relationships
+- Precise coordinate calculations (x, y, z positions)
+- Automatic trajectory derivation from physics principles
+- Explicit spatial reasoning in every response
+
+### 2. **Physics Modeling** ⚙️
+- Proficient in classical mechanics, electromagnetism, thermodynamics
+- **API-Driven**: Generates PhysicsBridge API calls directly
+- Supports rigid body collisions, SPH fluids, ODE/PDE solvers
+- Physics principles → API translation in reasoning chain
+
+### 3. **Advanced Problem Solving** 🏆
+- Optimized for CPhO (Chinese Physics Olympiad) level problems
+- Optimized for IMO (International Mathematical Olympiad) level problems
+- Handles complex multi-body systems with constraints
+- Deep reasoning for competition-level problems
+
+## Training Details
+
+- **Base Model**: OLMo-3.1-32B-Instruct
+- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
+- **LoRA Rank**: 256
+- **Training Data**: 3,500+ physics and mathematics problems with detailed reasoning chains
+- **Training Framework**: DeepSpeed ZeRO-3 with BF16 precision
+- **Hardware**: 3x RTX 5090 GPUs
+
+## Model Output Format
+
+The model generates responses with **explicit spatial reasoning chains** that enforce coordinate thinking:
+
+```
+<thought>
+【空间直觉分析】
+- 问题的3D空间结构
+- 物体初始位置 (x₀, y₀, z₀)
+- 物体初始速度 (vₓ, vᵧ, vᵤ)
+- 坐标系建立
+
+【物理原理推导】
+- 适用的物理定律
+- 力的分析
+- 加速度计算
+
+【空间坐标计算】
+- 时间 t 时的位置: (x(t), y(t), z(t))
+- 速度向量: (vₓ(t), vᵧ(t), vᵤ(t))
+- 轨迹方程
+
+【物理→API翻译】
+- PhysicsBridge API 调用
+- 参数映射: 坐标 → API 参数
+- 初始条件设置
+</thought>
+
+<code>
+# PhysicsBridge API 调用
+physics = PhysicsBridge()
+physics.create_rigid_body(
+    position=(x₀, y₀, z₀),
+    velocity=(vₓ, vᵧ, vᵤ),
+    mass=m,
+    shape='sphere'
+)
+# ... 更多 API 调用
+</code>
+```
+
+### What Makes This Different
+
+✨ **Forced Spatial Thinking**: The model MUST calculate coordinates before writing code
+✨ **API-First Architecture**: Physics principles are translated directly to PhysicsBridge API
+✨ **Explicit Reasoning**: Every step of spatial reasoning is visible in the output
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "SStarrySSky/Askit-OLMo-32B-Spatial-Thinking-Preview"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
+
+# Example: Physics simulation with spatial reasoning
+prompt = """
+Create a physics simulation for a ball dropped from 10 meters.
+The ball has mass 1kg and initial velocity (0, 0, 0).
+Use PhysicsBridge API to create the simulation.
+"""
+
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_length=2048, temperature=0.7)
+response = tokenizer.decode(outputs[0])
+
+# Output will include:
+# 1. <thought> section with explicit spatial calculations
+# 2. <code> section with PhysicsBridge API calls
+print(response)
+```
+
+## Use Cases
+
+### 1. **Physics Education** 🎓
+- Generate interactive animations for teaching physics concepts
+- Explicit spatial reasoning helps students understand 3D motion
+- API-driven code is directly executable in Askit platform
+
+### 2. **Mathematical Visualization** 📊
+- Create visual demonstrations of mathematical problems
+- Coordinate calculations ensure geometric accuracy
+- Perfect for IMO-level problem visualization
+
+### 3. **Research Simulation** 🔬
+- Build physics simulations for academic research
+- Spatial thinking ensures correct coordinate systems
+- API integration with PhysicsBridge for real-time rendering
+
+### 4. **Competitive Problem Solving** 🏆
+- Solve CPhO and IMO level problems with visualization
+- Forced spatial reasoning matches competition requirements
+- Generates production-ready simulation code
+
+## Limitations
+
+- Optimized for physics and mathematics domains
+- Requires PhysicsBridge API for full functionality
+- Best performance with detailed problem descriptions
+
+## License
+
+GPL-3.0 License
+
+## Citation
+
+```bibtex
+@model{askit2024,
+  title={Askit-OLMo-32B-Spatial-Thinking-Preview},
+  author={Starry Sky},
+  year={2024}
+}
+```
+
+## Acknowledgments
+
+Built on top of OLMo-3.1-32B-Instruct by Allen Institute for AI.