Downloads · 30 days
0
AIME-God-v5.2-Elite-1.5B
yesaf990/AIME-God-v5.2-Elite-1.5B
AIME-God-v5.2-Elite-1.5B is a machine learning model from yesaf990. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
AIME-God v5.2 "Elite" is a specialized reasoning model fine-tuned for high-level competitive mathematics (AIME, Olympiad, AMC). It was trained on the "Elite Turbo" dataset—a curated collection of 28,500 high-density r…
Access
Public
Updated Mar 20, 2026
Repo size
85.3 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors73.9 MB · 87%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- Mar 20, 2026
- Updated
- Mar 20, 2026
- SHA
- 5406271a
Base models
Model card
From the Hugging Face model README
🔱 AIME-God v5.2 "Elite" (1.5B Reasoning)
AIME-God v5.2 "Elite" is a specialized reasoning model fine-tuned for high-level competitive mathematics (AIME, Olympiad, AMC). It was trained on the "Elite Turbo" dataset—a curated collection of 28,500 high-density reasoning traces designed to master the most challenging contest problems.
👤 Credits: Abdullah Fowzan
This model and the underlying "God-Mode" dataset generation architecture were developed by Abdullah Fowzan.
🚀 Model Highlights
- Architecture: Qwen 2.5 1.5B-Instruct (4-bit QLoRA)
- Specialization: AIME Problems 11-15 and Tier-15 Olympiad methods.
- Training Strategy: 1.5B instruction-tuned base model + Elite reasoning traces (28.5k samples).
- Format: Zero-System Prompt optimized (merged instructions into User turn).
🧠 Tier-15 Mastery Documentation
The model is specifically trained on advanced mathematical lemmas and techniques:
- LTE Lemma (Lifting The Exponent): For advanced valuations in Number Theory.
- Geometric Inversion: Circle/Line inversion for hard geometry.
- Barycentric Coordinates: Professional triangle solving.
- Generating Functions: Specialized combinatorics and partition theory.
- Roots of Unity Filter: Complex number summation for algebra.
📊 Dataset Composition
| Source | Count | Description |
|---|---|---|
| Custom Reasoning traces | 10,000 | Hand-crafted traces using Tier-15 methods. |
| Competition Math | 10,000 | Sampled from AIME, Numina, and MATH (Elite subsets). |
| Synthetic God-Mode | 8,500 | Machine-generated chains of thought for difficult logic. |
🛠️ Usage
It is recommended to use the model with a temperature between 0.0 and 0.3 for maximum deterministic accuracy in mathematical steps.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "yesaf990/AIME-God-v5.2-Elite-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
# Use with standard Qwen 2.5 prompt template
Developed for the Benchmark-Crusher Initiative. 🚀💎🔱
- License
- apache-2.0
- Languages
- en
- Created
- Mar 20, 2026
- Updated
- Mar 20, 2026