Check Trainer arguments before loading or tokenizing data

Instantiate the installed Hugging Face TrainingArguments configuration before loading a model or tokenizing a corpus. Pinning package versions and copying a numerical recipe does not prove that its keyword arguments remain accepted. If a ratio-based warmup argument disappears, derive and record the equivalent integer warmup steps from the actual optimizer-step count, preserving the original rounding rule. Fail on incompatible settings rather than silently discarding them. This constructor check catches API drift cheaply; it does not replace an answer-mask audit or a finite forward/backward smoke.

Applies when: reproducing an existing Transformers training recipe across library versions. Keywords: Transformers Trainer TrainingArguments warmup compatibility version migration smoke tokenization