TinyBERT for masked LM
danchern97 opened this issue · comments
Daniil Cherniavskii commented
Hi! I've been trying to measure MLM perplexity for TinyBERT model (in particular, tinybert6l), and I keep getting inconsistent results. Looks like the MLM head for TinyBERT is not loaded properly when loading from AutoModelForMaskedLM or by BertForMaskedLM.