ModernBERT-base-zeroshot-v2.0

Model description

This model is answerdotai/ModernBERT-base fine-tuned on the same dataset mix as the zeroshot-v2.0 models in the Zeroshot Classifiers Collection.

The model is very fast and memory efficient. It's multiple times faster and consumes multiple times less memory than DeBERTav3. The memory efficiency enables larger batch sizes. I got a ~2x speed increase by enabling bf16 (instead of fp16).
It performs slightly worse then DeBERTav3 on average on the tasks tested below.
I'm in the process of preparing a newer version trained on better synthetic data to make full use of the 8k context window and to update the training mix of the older zeroshot-v2.0 models.

Per-dataset breakdown:

Datasets	Mean	Mean w/o NLI	mnli_m	mnli_mm	fevernli	anli_r1	anli_r2	anli_r3	wanli	lingnli	wellformedquery	rottentomatoes	amazonpolarity	imdb	yelpreviews	hatexplain	massive	banking77	emotiondair	emocontext	empathetic	agnews	yahootopics	biasframes_sex	biasframes_offensive	biasframes_intent	financialphrasebank	appreviews	hateoffensive	trueteacher	spam	wikitoxic_toxicaggregated	wikitoxic_obscene	wikitoxic_identityhate	wikitoxic_threat	wikitoxic_insult	manifesto	capsotu
Accuracy	0.831	0.835	0.932	0.936	0.884	0.763	0.647	0.657	0.823	0.889	0.753	0.864	0.949	0.935	0.974	0.798	0.788	0.727	0.789	0.793	0.489	0.893	0.717	0.927	0.851	0.859	0.907	0.952	0.926	0.726	0.978	0.912	0.914	0.93	0.951	0.906	0.476	0.708
F1 macro	0.813	0.818	0.925	0.93	0.872	0.74	0.61	0.611	0.81	0.874	0.751	0.864	0.949	0.935	0.974	0.751	0.738	0.746	0.733	0.798	0.475	0.893	0.712	0.919	0.851	0.859	0.892	0.952	0.847	0.721	0.966	0.912	0.914	0.93	0.942	0.906	0.329	0.637
Inference text/sec (A100 40GB GPU, batch=128)	3472.0	3474.0	2338.0	4416.0	2993.0	2959.0	2904.0	3003.0	4647.0	4486.0	5032.0	4354.0	2466.0	1140.0	1582.0	4392.0	5446.0	5296.0	4904.0	4787.0	2251.0	4042.0	1884.0	4048.0	4032.0	4121.0	4275.0	3746.0	4485.0	1114.0	4322.0	2260.0	2274.0	2189.0	2085.0	2410.0	3933.0	4388.0

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 32
eval_batch_size: 128
seed: 42
optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.06
num_epochs: 2