biolinkbert-mednli

This model is a fine-tuned version of michiyasunaga/BioLinkBERT-large on MedNLI. It achieves the following results on the evaluation set:

{
    "eval_accuracy": 0.8788530230522156,
    "eval_loss": 0.7843484878540039,
    "eval_runtime": 39.7009,
    "eval_samples": 1395,
    "eval_samples_per_second": 35.138,
    "eval_steps_per_second": 1.108
}

The accuracy for the test set is

{
    "eval_accuracy": 0.8607594966888428,
    "eval_loss": 0.879707932472229,
    "eval_runtime": 27.4404,
    "eval_samples": 1395,
    "eval_samples_per_second": 51.821,
    "eval_steps_per_second": 1.64
}

The labels are

"id2label": {
    "0": "entailment",
    "1": "neutral",
    "2": "contradiction"
  },

Training procedure

This model checkpoint is made by mednli.py by the following command:

root=/path/to/mednli/;
python mednli.py \
    --model_name_or_path michiyasunaga/BioLinkBERT-large \
    --do_train --train_file ${root}/mli_train_v1.jsonl \
    --do_eval --validation_file ${root}/mli_dev_v1.jsonl \
    --do_predict --test_file ${root}/mli_test_v1.jsonl \
    --max_seq_length 512 --fp16 --per_device_train_batch_size 16 --gradient_accumulation_steps 2 \
    --learning_rate 3e-5 --warmup_ratio 0.5 --num_train_epochs 10 \
    --output_dir ./biolinkbert_mednli

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.5
• num_epochs: 10.0
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.22.2
• Pytorch 1.13.0+cu117
• Datasets 2.4.0
• Tokenizers 0.12.1