Rollback to enhanced version with YAML metadata

7dd5fd9 verified about 1 month ago

7.16 kB

language:
  - en
  - zh
license: gpl-3.0
tags:
  - physics
  - mathematics
  - spatial-reasoning
  - code-generation
  - ollama
  - gguf
  - lora
  - deepspeed
  - qlora
  - fine-tuning
datasets:
  - physics-problems
  - mathematics-problems
model-index:
  - name: Askit-OLMo-32B-Spatial-Thinking-Preview
    results:
      - task:
          type: text-generation
        dataset:
          type: physics-math-problems
          name: CPhO & IMO Level Problems
        metrics:
          - type: accuracy
            value: High spatial reasoning accuracy

🚀 Askit-OLMo-32B-Spatial-Thinking-Preview

AI-Powered Physics Simulation & Mathematical Animation Generation

Explicit Spatial Reasoning + API-Centric Code Generation

✨ What Makes Askit Special?

Askit-OLMo-32B-Spatial-Thinking-Preview is not just another code generation model. It's a spatial reasoning specialist that thinks in 3D coordinates before writing code.

🧠 Core Innovation: Forced Spatial Cognition

Unlike traditional models that generate code directly, Askit enforces a complete spatial thinking pipeline:

Physics Problem
    ↓
【空间直觉分析】 (3D Space Analysis)
    ↓
【空间坐标计算】 (Coordinate Calculations)
    ↓
【物理→API翻译】 (Physics → API Translation)
    ↓
Production-Ready Code

Key Features:

✅ Forced Spatial Cognition: Every problem decomposed into 3D coordinates
✅ API-Centric Reasoning: Physics principles → PhysicsBridge API calls
✅ Explicit Reasoning: Complete thinking chain visible in output
✅ Competition-Ready: Optimized for CPhO & IMO level problems

🎯 Capabilities

1. Spatial Understanding 🗺️

3D object position relationships
Precise coordinate calculations (x, y, z)
Automatic trajectory derivation
Explicit spatial reasoning in every response

2. Physics Modeling ⚙️

Classical mechanics, electromagnetism, thermodynamics
API-Driven: Direct PhysicsBridge API generation
Rigid body collisions, SPH fluids, ODE/PDE solvers
Complex multi-body systems with constraints

3. Advanced Problem Solving 🏆

CPhO (Chinese Physics Olympiad) level optimization
IMO (International Mathematical Olympiad) level optimization
Deep reasoning for competition problems
Production-ready simulation code

🛠️ Ecosystem & Integration

📱 Official Software

Askit. - Interactive Physics Animation Platform

🌐 Website: askit.space
💻 GitHub: github.com/SStarrySSky/Askit.
📦 Features:
- Real-time physics simulation with OpenGL rendering
- Interactive controls (sliders, buttons, checkboxes)
- Timeline-based animation control
- Snapshot system for AI context awareness
- PhysicsBridge API integration

🔗 Integration Points

Askit-OLMo-32B Model
        ↓
    Generates Code
        ↓
PhysicsBridge API
        ↓
Askit. Platform
        ↓
Real-time Visualization

📚 Related Projects

Project	Purpose	Link
Askit.	Interactive Animation Platform	GitHub
PhysicsBridge	Physics Engine Wrapper	Integrated in Askit.
OLMo-3.1-32B	Base Model	Allen AI

📊 Model Specifications

Aspect	Details
Base Model	OLMo-3.1-32B-Instruct
Fine-tuning	LoRA (Rank 256)
Training Data	3,500+ physics/math problems
Framework	DeepSpeed ZeRO-3 + BF16
Hardware	3x RTX 5090 GPUs
Output Format	Explicit reasoning chains + code

💡 Output Format

The model generates complete reasoning chains with explicit spatial thinking:

<thought>
【空间直觉分析】
- 3D space structure analysis
- Initial positions (x₀, y₀, z₀)
- Initial velocities (vₓ, vᵧ, vᵤ)
- Coordinate system setup

【物理原理推导】
- Applicable physics laws
- Force analysis
- Acceleration calculations

【空间坐标计算】
- Position at time t: (x(t), y(t), z(t))
- Velocity vector: (vₓ(t), vᵧ(t), vᵤ(t))
- Trajectory equations

【物理→API翻译】
- PhysicsBridge API calls
- Parameter mapping: coordinates → API
- Initial conditions setup
</thought>

<code>
# PhysicsBridge API Integration
physics = PhysicsBridge()
physics.create_rigid_body(
    position=(x₀, y₀, z₀),
    velocity=(vₓ, vᵧ, vᵤ),
    mass=m,
    shape='sphere'
)
# ... more API calls
</code>

🚀 Quick Start

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")

# Physics simulation with spatial reasoning
prompt = """
Create a physics simulation for a ball dropped from 10 meters.
Ball mass: 1kg, initial velocity: (0, 0, 0)
Use PhysicsBridge API.
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=2048, temperature=0.7)
print(tokenizer.decode(outputs[0]))

📖 Use Cases

🎓 Physics Education

Interactive animations for teaching concepts
Explicit spatial reasoning aids student understanding
API-driven code runs directly in Askit. platform

📊 Mathematical Visualization

Visual demonstrations of math problems
Geometric accuracy through coordinate calculations
Perfect for IMO-level problem visualization

🔬 Research Simulation

Academic research physics simulations
Correct coordinate systems guaranteed
Real-time rendering via PhysicsBridge

🏆 Competitive Problem Solving

CPhO and IMO level problem solving
Forced spatial reasoning matches competition requirements
Production-ready simulation code

🔗 Links & Resources

Official Channels

🌐 Website: askit.space
💻 GitHub Repository: github.com/SStarrySSky/Askit.
🤗 HuggingFace Model: SStarrySSky/Askit-OLMo-32B-Spatial-Thinking-Preview

Base Technologies

🧠 OLMo-3.1-32B: allenai/OLMo-3.1-32B-Instruct
⚙️ Transformers: huggingface.co/transformers

📄 License

GPL-3.0 License - See LICENSE

🙏 Acknowledgments

Built on top of OLMo-3.1-32B-Instruct by Allen Institute for AI.

Integrated with Askit. - Interactive Physics Animation Platform.

Made with ❤️ by Starry Sky

🌐 Website • 💻 GitHub • 🤗 HuggingFace