New README.md file

README.md

@@ -1,11 +1,92 @@
+# AMC12 Dataset (Research-Oriented)
+
+A structured dataset derived from the AMC 12 (American Mathematics Competitions), designed for **LLM training, evaluation, and reinforcement learning (RL)** on mathematical reasoning tasks.
+
+This repository contains **all AMC 12 problems from 2000–2025**, making it one of the most complete AMC12 datasets available for research.
+
 ---
-task_categories:
-- multiple-choice
-- reinforcement-learning
-language:
-- en
-tags:
-- agent
-size_categories:
-- 1K<n<10K
----
+
+## 📘 Introduction
+
+The AMC 12 is a **25-question, 75-minute multiple-choice examination** aimed at high school students. Problems are designed to **increase in difficulty progressively**, requiring a combination of algebra, geometry, combinatorics, and number theory reasoning.
+
+* **Format:** 25 multiple-choice questions (A–E)
+* **Duration:** 75 minutes
+* **Difficulty progression:** Problems 1 → 25 increase in complexity
+* **Calculator policy:**
+
+  * Since 2008, calculators are **not permitted**
+  * Problems are designed to be solvable without computational aids
+
+Top-performing students (~top 6%) are invited to participate in the AIME, making AMC 12 a strong proxy for **high-level mathematical reasoning ability**.
+
+---
+
+## 📦 Dataset Overview
+
+Each sample corresponds to a single AMC 12 problem instance, represented in a **structured, model-ready format**.
+
+### Difficulty Structure
+
+We adopt a coarse-grained difficulty approximation aligned with problem order:
+
+| Problem Range | Difficulty      |
+| ------------- | --------------- |
+| 1–10          | Easy–Medium (2) |
+| 11–20         | Medium–Hard (3) |
+| 21–25         | Hard (4)        |
+
+This structure enables:
+
+* Curriculum learning
+* Difficulty-aware evaluation
+* Model capability stratification
+
+---
+
+## 🚀 Why This Dataset?
+
+### Compared to Other Math Datasets
+
+* **Not heavily pretrained**
+  Unlike datasets such as GSM8K, AMC-style problems are less likely to be memorized by models
+
+* **Higher reasoning complexity**
+  Problems typically require **multi-step, structured reasoning**, often exceeding datasets like MATH500
+
+* **Clean evaluation signal**
+
+  * Multiple-choice format eliminates ambiguity
+  * No unit mismatch issues (e.g., “8 months vs 240 days”)
+
+* **Fully verifiable**
+  Every problem has a **unique, discrete answer**, ideal for RL reward design
+
+---
+
+### Compared to Other AMC Datasets
+
+* **Complete coverage (2000–2025)**
+  Includes all AMC12A and AMC12B problems across 25 years
+
+* **Fully indexed & traceable**
+  Each problem maps directly to its original contest and position
+
+* **Structured for ML pipelines**
+  Ready for:
+
+  * RL training (PPO / GRPO)
+  * Pass@k evaluation
+  * Verifier-based reward systems
+
+---
+
+## 📚 Data Source & Attribution
+
+This dataset is curated from publicly available resources, with primary reference to:
+
+* Art of Problem Solving
+
+All AMC problems are **copyrighted by the Mathematical Association of America (MAA)** under the American Mathematics Competitions program.
+
+This repository does **not claim ownership** of the original problem statements and provides them solely for research and educational purposes.