ReasonForge

Deterministic math tools for small language models.

ReasonForge gives small LLMs (8B–32B) access to a verified SymPy computation backend via tool calling. Instead of relying on the model to compute, all math is delegated to deterministic tools — the model only reasons about what to compute and how to present results.

Architecture

User Question → LLM (Qwen3) → Tool Calls → SymPy Backend → Verified Results → LLM → Final Answer

Two-phase response pipeline:

1. Compute (/no_think): Model calls tools with thinking disabled — forced delegation
2. Present (thinking ON): Model reasons about verified results, composes the answer

Tools

Tool	Operations	Backend
math_tool	compute, solve, simplify, factor, expand, gcd, lcm, prime_factors, divisors, mod_inverse, nsolve, crt + SymPy builtins (totient, fibonacci, isprime...)	SymPy
calculus_tool	differentiate, integrate, limit, series, summation, partial_fraction, trigsimp, ode_solve, laplace	SymPy
matrix_tool	determinant, inverse, eigenvalues, eigenvectors, rank, rref, transpose, multiply, add, trace, nullspace, columnspace, charpoly, norm, adjugate, solve (Ax=b)	SymPy
statistics_tool	describe, mean, median, mode, std, variance, correlation, regression, percentile, zscore, skewness, kurtosis, geometric_mean, harmonic_mean	Python stdlib

Project Structure

MCP/
├── experts/math/
│   ├── server.py              # MCP server entry point
│   └── tools/
│       ├── preprocess.py      # Expression parser (^ → **, implicit multiplication, infinity handling)
│       ├── algebra.py         # math_tool — algebra + number theory
│       ├── calculus.py        # calculus_tool — derivatives, integrals, ODEs, etc.
│       ├── matrix.py          # matrix_tool — linear algebra
│       └── statistics.py      # statistics_tool — descriptive & inferential stats
├── tests/
│   └── test_math_tools.py     # Unit tests for all tools
├── ui/
│   ├── app.py                 # Gradio chat interface with two-phase pipeline
│   └── style.css              # Custom UI styles (dark mode, input pill, etc.)
├── ReasonForge_Colab.ipynb    # One-click Colab deployment notebook
├── pyproject.toml
└── run_ui.bat                 # Local launcher (Windows)

Quick Start (Local)

# Requires: Ollama running with qwen3:8b
uv sync
uv run python -m ui.app
# Opens at http://localhost:7861

Colab Deployment (GPU)

Open ReasonForge_Colab.ipynb in Google Colab Pro with an A100 GPU. It clones this repo, installs Ollama + qwen3:32b, and launches the UI with a public Gradio link.

Running Tests

uv run python -m tests.test_math_tools

Tech Stack

• LLM Backend: Ollama (local) or any OpenAI-compatible API
• Math Engine: SymPy — symbolic computation
• UI: Gradio — chat interface with LaTeX rendering
• Protocol: MCP (Model Context Protocol) compatible