evaluate — HuggingFace Evaluate 0.4.6 (iOS)

Bundled and device-verified. Computes metrics for training/eval loops.

What works

import evaluate and local metric compute
evaluate.load("accuracy") → .compute(references=..., predictions=...) (device-verified → {'accuracy': 0.75})
Combining metrics, evaluate.combine([...])

Feature	Status	Workaround
`evaluate.load("<name>")` first call	Downloads the metric script from the Hub	Needs network once; the script then runs fully offline. Or compute the metric inline (accuracy/F1 are a few lines)
Metrics needing extra C/Rust deps (e.g. `sacrebleu`, `bleurt`)	The metric script imports a dep that may not be bundled	Stick to pure-Python metrics (accuracy, f1, precision, recall, mse, mae)