Update README.md

README.md

@@ -1,3 +1,47 @@
----
-license: mit
----
+---
+license: mit
+language:
+- en
+base_model:
+- meta-llama/Llama-3.2-3B-Instruct
+---
+
+# Gladiator-Mini-exp-1211: A Compact and Powerful Reasoning Engine
+
+**Gladiator-Mini-exp-1211** is a 3-billion parameter language model designed for **complex reasoning tasks**. This experimental model, based on [meta-llama/Llama-3.2-3B-Instruct](link to the base model), offers surprisingly strong analytical capabilities for its size. It demonstrates the potential of smaller models to achieve impressive performance in analytical thinking. We chose to finetune on a Llama model due to finetuning difficulties with the Qwen 2.5 4B model.
+
+**What Makes it Stand Out?**
+
+This model is focused on **reasoning**. It was trained on a specialized dataset emphasizing **chain-of-thought reasoning, multi-step problem-solving, and the ability to consider problems from multiple perspectives**. It shows particular promise in **mathematics**. Although still in development, Gladiator-Mini-exp-1211 has demonstrated the ability to solve problems that often challenge larger models. For instance, it can correctly count the number of "r's" in "strawberry" when given an appropriate system prompt - a task that many larger models struggle with. It can also accurately solve the classic "farmer's dilemma" problem, determining the number of chickens and pigs based on the total number of heads and legs. The model can even correctly apply the formula for an infinite geometric series. However, it can still be susceptible to subtle mathematical misconceptions, like the common misunderstanding that 0.999... is not exactly equal to 1. This highlights the need for ongoing refinement and careful evaluation of its reasoning.
+
+**How to Use It (Important!)**
+
+Gladiator-Mini-exp-1211 works best with **carefully crafted system prompts** that guide its reasoning. These prompts are essential to unlocking its analytical abilities. Without a suitable prompt, its performance may be compromised, **and it could even produce incorrect solutions**. We are working to reduce its reliance on these prompts in the future.
+
+**Here are some example system prompts:**
+
+*   **For General Reasoning:** "You are a highly advanced AI with deep reasoning capabilities. You are able to explain your thinking process step by step, to ensure full accuracy."
+*   **For Mathematical Problems:** "You are a highly advanced mathematical reasoning AI, capable of complex chain-of-thought processing and multi-step problem-solving. You should approach each problem meticulously, breaking it down into smaller, manageable steps, explaining your reasoning clearly and concisely."
+*   **For Logical Deduction:** "You are a master logician AI, capable of dissecting complex scenarios and identifying underlying principles. You excel at deductive reasoning, pattern recognition, and articulating your thought process with precision."
+
+**Why is it "Experimental"?**
+
+Gladiator-Mini-exp-1211 is still under development (hence the "exp" in its name). It may have biases, limitations, and areas where its reasoning is not yet perfect. We're releasing it to the community to gather feedback, explore its full capabilities, and continue to improve it. The experimental date is 12/11/2024.
+
+**The Goal:**
+
+Our objective is to create a powerful reasoning engine in a compact package. We believe that smaller, more efficient models are the future of AI, and Gladiator-Mini-exp-1211 represents a significant step in that direction. We hope that it will eventually rival, or even surpass, the performance of larger models on reasoning-intensive tasks, while remaining far more resource-efficient. We even hope it can beat the o1 model series in complex reasoning tasks.
+
+**Contribute to the Experiment!**
+
+We encourage you to experiment with Gladiator-Mini-exp-1211, test its limits, and share your findings. Together, we can push the boundaries of what's possible with smaller, more efficient AI models.
+
+**Limitations:**
+
+*   Requires system prompts for optimal performance. An incorrect system prompt can lead to an incorrect answer, while a correct prompt usually leads to a correct one.
+*   Reasoning abilities are still being refined.
+*   May exhibit biases or inaccuracies.
+
+**Disclaimer:**
+
+Gladiator-Mini-exp-1211 is an experimental model and should be used with caution. Please critically evaluate its outputs and do not take them as absolute truth.