VLQM-1.5B-Coder

🚀 VLQM-1.5B-Coder

A fine-tuned language model for generating Manim animation code from natural language descriptions.

---

📖 Model Description

VLQM-1.5B-Coder is a specialized language model fine-tuned to generate Manim Python code from natural language descriptions. Manim is the mathematical animation engine famously used by 3Blue1Brown.

What This Model Does

Given a natural language prompt:

"Create a blue circle that moves to the right"

The model generates valid Manim Python code:

from manim import *

class GenScene(Scene):
    def construct(self):
        circle = Circle(color=BLUE)
        self.add(circle)
        self.play(circle.animate.shift(RIGHT * 3))

---

🎯 Intended Use

• Primary Use: Generating Manim animation code from text descriptions
• Users: Developers, educators, content creators making math/science animations
• Languages: English prompts, Python code output

Example Use Cases

• Creating educational math animations
• Generating visualizations for presentations
• Prototyping Manim scenes quickly
• Learning Manim syntax through examples

---

📊 Training Details

Hyperparameters

---

---

📈 Model Performance

Training Loss Curve

The model shows a strong convergence pattern, with validation loss stabilizing around 0.71 and training loss reaching 0.48.

⚠️ Limitations

Known Limitations

1. Complex Animations: May struggle with multi-step animations involving many objects
2. Advanced Manim Features: Less reliable with 3D scenes, complex graphs, or advanced camera movements
3. API Hallucinations: Sometimes generates non-existent Manim methods (e.g., axes.get_sine())
4. Indentation Issues: Occasionally produces incorrectly indented code
5. Long Prompts: Performance degrades with very long or complex descriptions

What This Model is NOT

• ❌ Not a general-purpose code generator
• ❌ Not trained for non-Manim Python code
• ❌ Not suitable for production without human review
• ❌ Not a replacement for learning Manim fundamentals

Recommended Practices

• ✅ Always review and test generated code before use
• ✅ Use simple, clear prompts for best results
• ✅ Keep prompts focused on one animation at a time
• ✅ Be prepared to make minor edits to fix issues

---

🚀 Quick Start

Using with Transformers (Cross-Platform)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "vikramlingam/VLQM-1.5B-Coder",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("vikramlingam/VLQM-1.5B-Coder")

prompt = "Create a red square that rotates"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Using with MLX (Apple Silicon)

python -m mlx_lm.generate \
    --model vikramlingam/VLQM-1.5B-Coder \
    --prompt "Create a circle animation" \
    --max-tokens 256

---

📁 Model Files

VLQM-1.5B-Coder/
├── config.json              # Model configuration
├── model.safetensors        # Model weights (~3 GB)
├── tokenizer.json           # Tokenizer
├── tokenizer_config.json
├── vocab.json
├── merges.txt
├── generation_config.json
├── special_tokens_map.json
└── README.md                # This file

---

📜 License

This model is released under the Apache 2.0 License.

• ✅ Commercial use allowed
• ✅ Modification allowed
• ✅ Distribution allowed
• ⚠️ Must include license and copyright notice

---

🙏 Acknowledgments

• Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct by Alibaba
• Dataset: generaleoley/manim-codegen
• Training Framework: Apple MLX
• Animation Engine: Manim Community

---

📬 Citation

If you use this model, please cite:

@misc{vlqm-1.5b-coder,
  author = {Vikram Lingam},
  title = {VLQM-1.5B-Coder: Manim Code Generation Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/vikramlingam/VLQM-1.5B-Coder}
}

---

📊 Model Card Contact

For questions or issues, please open a discussion on the model's Hugging Face page.