Firstly, the base-model's tokenizer was adapted and extended to better represent biomedical domain terms and reduce fragmentation.
Then, the adapted tokenizer was used in a continue-pretraining stage using the unlabelled task-specific corpus on Task-Adaptive finetuning(TAPT).
Finally,the resulting tapt-model was fine-tuned on NER task using the labelled version of the data from step 2 above.

This model uses the following label space:

The label schema determines the classifier output dimension and the evaluation setup used for this model.

This model is based on:

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 16
eval_batch_size: 16
seed: 3407
gradient_accumulation_steps: 2
total_train_batch_size: 32
optimizer: Use OptimizerNames.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: 0.01
num_epochs: 20
mixed_precision_training: Native AMP

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
0.515	0.3802	100	0.1174	0.3626	0.3220	0.3411	0.9563
0.1112	0.7605	200	0.0929	0.4919	0.3718	0.4235	0.9667
0.0736	1.1407	300	0.0747	0.5468	0.5980	0.5712	0.9737
0.055	1.5209	400	0.0680	0.6491	0.6522	0.6507	0.9776
0.0481	1.9011	500	0.0681	0.6204	0.7192	0.6661	0.9731
0.0383	2.2814	600	0.0722	0.6468	0.6448	0.6458	0.9738
0.0339	2.6616	700	0.0709	0.6307	0.6597	0.6449	0.9726
0.0332	3.0418	800	0.0656	0.7026	0.6986	0.7006	0.9789
0.0272	3.4221	900	0.0732	0.7110	0.6693	0.6895	0.9785
0.026	3.8023	1000	0.0790	0.7586	0.6544	0.7027	0.9787
0.0234	4.1825	1100	0.0703	0.6960	0.7332	0.7141	0.9798
0.0209	4.5627	1200	0.0773	0.6497	0.6759	0.6625	0.9770
0.0216	4.9430	1300	0.0742	0.6638	0.7550	0.7065	0.9776
0.0172	5.3232	1400	0.0804	0.7031	0.6732	0.6878	0.9780