Lense - Loss Curve Analyzer

A lightweight Python tool to visualize and analyze machine learning training loss curves from JSON/JSONL logs, with built-in smoothing, metrics calculation, and outlier detection.

Features

• Parses JSON, JSONL, and single-quoted Python dict logs
• Robust smoothing (SciPy or NumPy fallback moving average)
• Calculates key metrics: slope, SNR, best loss, variance, exposure
• Outlier detection using Median Absolute Deviation (MAD)
• Interactive Tkinter GUI with plot export and clipboard support
• CLI mode for headless use

Dependencies

• Python 3.8+
• Core: matplotlib, numpy
• Optional: scipy (faster smoothing), pyperclip (clipboard copy)

Install dependencies:

pip install -r requirements.txt

Usage

GUI Mode (default)

python lense.py

1. Paste or load JSONL logs into the input box
2. Click "Analyze Now" or press Ctrl+Enter
3. View metrics and interactive loss plot
4. Export metrics to clipboard or Markdown

CLI Mode

python lense.py path/to/training_logs.jsonl

Supported Log Format

Logs must contain a loss key (required), with optional fields like epoch, learning_rate, etc.:

{"loss": 0.52, "epoch": 1.0, "step": 100}
{"loss": 0.48, "epoch": 1.1, "step": 110}

Improvements & Fixes (2026-04-26)

• Replaced naive single-quote regex with ast.literal_eval for safer log parsing
• Removed unused calculate_grade dead code
• Fixed outlier detection to use MAD instead of standard deviation
• Corrected epoch boundary detection to trigger on whole epoch transitions
• Added requirements.txt and this README