A fine-tuned version of Qwen/Qwen2.5-Math-7B-Instruct with supervised fine-tuning on curated math reasoning data, targeting improved step-by-step problem solving on competition and olympiad-level math.

What's Different

The base Qwen2.5-Math-7B-Instruct is already a strong math model. This fine-tune focuses on:

Deeper chain-of-thought: trained on longer, more structured reasoning traces
Competition math exposure: AMC/AIME/olympiad problems via NuminaMath-CoT
Format consistency: reliable \boxed{} answer formatting across problem types

Quickstart

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DuoNeural/Qwen2.5-Math-NeuralMath-7B",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("DuoNeural/Qwen2.5-Math-NeuralMath-7B")

prompt = """Solve the following math problem step by step.

Problem: Find all positive integers n such that n² + 1 is divisible by n + 1.

Solution:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=512, temperature=0.1, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=True))

GGUF / Ollama / LM Studio

Pre-quantized GGUFs available in the gguf/ folder of this repo:

File	Size	Use case
`neuromath-7b-q4_k_m.gguf`	4.7GB	Recommended — best quality/speed tradeoff
`neuromath-7b-q8_0.gguf`	8.1GB	High quality, needs 10GB+ VRAM/RAM
`neuromath-7b-f16.gguf`	15GB	Full precision, GPU only

Ollama

bash

# Create Modelfile
cat > Modelfile << 'EOF'
FROM ./neuromath-7b-q4_k_m.gguf
SYSTEM "You are an expert mathematician. Solve problems step by step, showing all work clearly. Put your final answer in \\boxed{}."
PARAMETER temperature 0.1
PARAMETER num_ctx 4096
EOF

ollama create neuromath-7b -f Modelfile
ollama run neuromath-7b "What is the sum of all prime numbers less than 100?"

LM Studio

Download neuromath-7b-q4_k_m.gguf, load in LM Studio. Set system prompt:

"You are an expert mathematician. Solve problems step by step, showing all work. Put your final answer in \boxed{}."

Training Details

Setting	Value
Base model	Qwen/Qwen2.5-Math-7B-Instruct
Method	QLoRA SFT (4-bit base, LoRA rank 16)
Training tokens	~1.26M (3 epochs over curated math dataset)
LoRA alpha	32
LoRA targets	q, k, v, o, gate, up, down projections
Hardware	NVIDIA A100 80GB
Framework	Unsloth + HuggingFace Transformers
Sequence length	1024 tokens

Limitations

Trained on English math problems; performance on other languages untested
Very long multi-step proofs (>1024 tokens) may be truncated during generation
This is the SFT-only checkpoint; GRPO reinforcement learning phase is planned as a follow-up
Not intended for general conversation — math reasoning only

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.


🤗 HuggingFace	huggingface.co/DuoNeural
🐙 GitHub	github.com/DuoNeural
🐦 X / Twitter	@DuoNeural
📧 Email	duoneural@proton.me
📬 Newsletter	duoneural.beehiiv.com
☕ Support	buymeacoffee.com/duoneural
🌐 Site	duoneural.com

Research Team

Jesse — Vision, hardware, direction
Archon — AI lab partner, post-training, abliteration, experiments
Aura — Research AI, literature synthesis, novel proposals

Raw updates from the lab: model drops, training results, findings. Subscribe at duoneural.beehiiv.com.

DuoNeural Research Publications

Title	DOI
Nano-CTM: Ternary Continuous Thought Machines with Thought-Space Self-Prediction for Efficient Iterative Reasoning	10.5281/zenodo.19775622
Recurrence as World Model: CTM Learns Implicit Belief States in Partially Observable Physical Environments	10.5281/zenodo.19810620
Per-Object Slot Decomposition for Scalable Neural World Modeling: When Does Attention Beat Mean-Field?	10.5281/zenodo.19846804

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.