Updated Readme file

README.md

@@ -0,0 +1,126 @@
+# **MiniLLM Geometry Engine**
+
+MiniLLM Geometry Engine is a fine-tuned version of the TinyLlama 1.1B
+language model, optimized for generating geometry-related functions for
+a geometry engine. Fine-tuned on a custom Geometry Chain-of-Thought
+(CoT) dataset, this model excels at producing accurate and efficient
+mathematical functions for geometric computations, stored in the Hugging
+Face .safetensors format for secure and efficient model loading.
+
+## **Model Overview**
+
+-   **Base Model**: TinyLlama 1.1B
+
+-   **Fine-Tuning Dataset**: Custom Geometry CoT dataset, including
+    > 2D/3D shapes, trigonometry, and coordinate geometry problems
+
+-   **Model Format**: .safetensors for secure and efficient weight
+    > storage
+
+-   **Intended Use**: Generating geometry functions for educational
+    > tools, CAD software, game development, or computational geometry
+
+-   **License**: MIT
+
+-   **Contact**: Contact via Hugging Face profile
+
+## **Important Note:**
+
+-   Use the system_prompt.txt for the System Prompt, this provides the
+    > accurate results.
+
+-   You can extract and edit the system prompt. I hope to add new
+    > functions later.
+
+## **Capabilities**
+
+MiniLLM Geometry Engine interprets geometry-related queries and outputs
+functional Python code or mathematical expressions. For example, given a
+prompt like \"Generate a function to calculate the area of a triangle,\"
+it produces executable code or formulas with clear reasoning, leveraging
+its CoT fine-tuning for logical accuracy. Supported tasks include area,
+volume, distance, and intersection calculations.
+
+## **Technical Details**
+
+-   **Architecture**: Transformer-based, inherited from TinyLlama 1.1B
+
+-   **Fine-Tuning Details**: Trained on a dataset of geometry problems
+    > with step-by-step solutions
+
+-   **Output Format**: Python code or pseudocode compatible with
+    > geometry engine APIs
+
+-   **Performance**: Improved accuracy on geometry tasks compared to the
+    > base TinyLlama model, with low latency
+
+-   **Model Storage**: Uses .safetensors for secure and efficient
+    > loading with Hugging Face\'s safetensors library
+
+## **Usage**
+
+To load and use the model with Hugging Face\'s Transformers library:
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from safetensors.torch import load_file
+
+\# Load tokenizer and model
+
+tokenizer =
+AutoTokenizer.from_pretrained(\"your-username/minillm-geometry-engine\")
+
+model =
+AutoModelForCausalLM.from_pretrained(\"your-username/minillm-geometry-engine\")
+
+\# Example prompt
+
+prompt = \"Draw a point A at 3,4\"
+
+inputs = tokenizer(prompt, return_tensors=\"pt\")
+
+outputs = model.generate(\*\*inputs, max_length=200)
+
+print(tokenizer.decode(outputs\[0\], skip_special_tokens=True))
+
+Install required libraries:
+
+pip install safetensors transformers
+
+## **Training Details**
+
+-   **Dataset**: Custom Geometry CoT dataset with 10,000 geometry
+    > problems and solutions
+
+-   **Training Procedure**: Fine-tuned for 3 epochs on a single NVIDIA
+    > A100 GPU
+
+-   **Hyperparameters**: Learning rate: 2e-5, batch size: 16
+
+-   **Hardware**: NVIDIA A100 GPU
+
+## **Limitations**
+
+-   Supports very few functions since this was an experimental model,
+    > maybe I will add other functions later.
+
+-   Limited to geometry-related tasks and may not generalize to other
+    > mathematical domains.
+
+-   Extensive use of strict System Prompting. I aim to eliminate that
+    > also.
+
+## **How to Contribute**
+
+Submit issues or pull requests via the Hugging Face repository.
+Contributions to the dataset or model improvements are welcome.
+
+## **Acknowledgments**
+
+Thanks to the TinyLlama team for the base model and Hugging Face for the
+safetensors and transformers libraries.
+
+MiniLLM Geometry Engine is a lightweight, efficient solution for
+developers and researchers needing reliable geometry function
+generation, with the security and performance benefits of the
+.safetensors format.