SymTorch Benchmark (Quick + Full)

A lightweight benchmark suite to test symbolic distillation with SymTorch + PySR on:

• x²
• sin(x) (noisy)
• x·exp(-x²)
• Normal PDF
• Black-Scholes call pricing (2D input)
• x·y (2D interaction)

What this project does

• Trains small MLPs on known target functions.
• Distills trained models into symbolic equations with SymTorch/PySR.
• Reports both neural-network and symbolic metrics.
• Writes a Markdown report for each run.

Requirements

• Python 3.11 recommended
• Julia (PySR may install/configure it automatically on first run)

Install packages:

pip install -r requirements.txt

Run

Quick mode (faster iteration):

python symtorch_benchmark.py --quick --report quick_benchmark_report.md

Full mode:

python symtorch_benchmark.py --report benchmark_report.md

If --report is omitted, an auto-named report is generated in the project root.

Output

• Markdown summary report in project root (e.g. quick_benchmark_report.md)
• PySR artifacts in SR_output/

Notes

• Quick mode reduces samples/epochs/iterations, while keeping learning meaningful.
• Training uses ReduceLROnPlateau and early stopping.
• Symbolic metrics are computed from symbolic model predictions (not placeholders).

Main files

• symtorch_benchmark.py — benchmark runner
• requirements.txt — Python dependencies
• symtorch_benchmark_guide.md — benchmark design notes
• quick_benchmark_report_v2.md — latest example report