Examples
Examples
Try the starting example
- Records: ID, predicted positive probability, binary actual label
- A,0.1,0 B,0.9,1 C,0.5,1
- Explicit symmetric probability clipping epsilon
- 0.01
- Mean binary log loss (natural-log units)
- 0.301289
- Per-record binary log loss
- A: 0.105360515658; B: 0.105360515658; C: 0.69314718056
- Number of probabilities changed by clipping
- 0
- Records evaluated
- 3
Try another scenario
- Records: ID, predicted positive probability, binary actual label
- A,0.1,0 B,0.9,1 C,0,1
- Explicit symmetric probability clipping epsilon
- 0.01
- Mean binary log loss (natural-log units)
- 1.605297
- Per-record binary log loss
- A: 0.105360515658; B: 0.105360515658; C: 4.60517018599
- Number of probabilities changed by clipping
- 1
- Records evaluated
- 3
Calculation method
Calculation method
Mean binary log loss (natural-log units)
Mean of −ln(p) for actual 1 and −ln(1−p) for actual 0, after stated clipping.
Per-record binary log loss
Each record loss; impossible unclipped endpoint is displayed as Infinite.
Number of probabilities changed by clipping
Count of supplied probabilities altered by max(epsilon,min(1−epsilon,p)).
Records evaluated
Number of supplied records.
Symbols and units
- a
- Records: ID, predicted positive probability, binary actual label
- b
- Explicit symmetric probability clipping epsilon
Use the units shown. A list keeps the order in which you enter its values. The symbol ^ means “raised to the power of”.
Assumptions & limits
Assumptions & limits
- Binary unweighted natural-log loss for supplied probabilities. Explicit epsilon controls clipping; epsilon 0 preserves impossible-outcome infinity. The displayed results do not run or retrain a model.
Are these inputs private?
These calculations run in your browser. A shared link contains the values you choose to share. Only questions submitted to the separate AI finder are sent to OpenAI.
Further reading
Background reading; this does not imply endorsement or independent review of SumFizz.