metadata
license: apache-2.0
language:
- en
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- gguf
library_name: transformers
pipeline_tag: text-generation
datasets:
- Rimyy/problemMath-Llama3.5K
base_model: unsloth/llama-3.2-3b-instruct-bnb-4bit
model_name: llama-3.2-3b-instruct-bnb-4bit-math-gguf
🧮 LLaMA 3.2 3B Instruct (Unsloth 4-bit) — Finetuned on Rimyy/problemMath-Llama3.5K (GGUF)
This model is a 4-bit GGUF variant of unsloth/llama-3.2-3b-instruct-bnb-4bit, fine-tuned on Rimyy/problemMath-Llama3.5K, a high-quality dataset of math reasoning and problem-solving questions. The model is tailored for math instruction, step-by-step reasoning, and educational applications.
🚨 Designed to reason, not just regurgitate. Small model, big brain.
🧠 Model Details
| Feature | Value |
|---|---|
| Base | unsloth/llama-3.2-3b-instruct-bnb-4bit |
| Finetuned Dataset | Rimyy/problemMath-Llama3.5K |
| Quantization | 4-bit GGUF (compatible with llama.cpp/text-generation-webui) |
| Format | GGUF |
| Language | English |
| Instruction Tuned | ✅ Yes |
📚 Dataset: Rimyy/problemMath-Llama3.5K
- ~3.5K math word problems and reasoning tasks
- Emphasizes chain-of-thought (CoT) explanations
- Covers arithmetic, algebra, and word problems
- Aligns with OpenAI-style "question → step-by-step answer" format
🔧 Quick Usage Example (llama.cpp)
./main -m llama-3.2-3b-math.gguf --prompt "### Question: What is the value of x if x + 3 = 7?
### Answer:"
Expected output:
To solve for x, subtract 3 from both sides of the equation:
x + 3 = 7
x = 7 - 3
x = 4
Answer: 4
🧪 Usage in Python
from llama_cpp import Llama
llm = Llama(
model_path="llama-3.2-3b-instruct-math.q4_K.gguf",
n_ctx=2048,
n_gpu_layers=32, # adjust based on your GPU
)
prompt = (
"### Question: If a rectangle has length 10 and width 5, what is its area?
"
"### Answer:"
)
response = llm(prompt)
print(response["choices"][0]["text"])
📦 Applications
- 🤖 Math tutoring agents
- 📚 AI-driven educational platforms
- 🧩 RAG pipelines for mathematical queries
- 📝 Automated solution generators
⚠️ Limitations
- Occasional step hallucinations
- Not optimized for LaTeX-heavy symbolic math
- May struggle on very long multi-step problems
📊 Qualitative Benchmark
| Task Type | Performance |
|---|---|
| Simple Arithmetic | ✅ Excellent |
| One-Step Algebra | ✅ Strong |
| Multi-Step CoT | ⚠️ Good (some drift) |
| Logic Puzzles | ⚠️ Mixed |
📌 Quantitative benchmarks forthcoming.
🔗 Citation
If you use this model, please cite:
@misc{rimyy2025math,
author = {Rimyy},
title = {ProblemMath-Llama3.5K: A Dataset for Math Problem Solving},
year = {2025},
url = {https://huggingface.co/datasets/Rimyy/problemMath-Llama3.5K}
}
🙌 Acknowledgements
- Meta for LLaMA 3.
- Unsloth for the 4-bit instruct base.
- Rimyy for an excellent math dataset.
- llama.cpp & GGUF community for stellar tooling.
🔢 Small enough to run on your laptop, smart enough to teach algebra.