Edit model card

distilbert-legal-chunk

This model is a fine-tuned version of distilbert-base-cased on the None dataset. It achieves the following results on the evaluation set:

  • Loss: 0.0699
  • Precision: 0.8994
  • Recall: 0.8721
  • Macro F1: 0.8855
  • Micro F1: 0.8855
  • Accuracy: 0.9789
  • Marker F1: 0.9804
  • Marker Precision: 0.9687
  • Marker Recall: 0.9925
  • Reference F1: 0.9791
  • Reference Precision: 0.9804
  • Reference Recall: 0.9778
  • Term F1: 0.8670
  • Term Precision: 0.8844
  • Term Recall: 0.8502

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-05
  • train_batch_size: 32
  • eval_batch_size: 32
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 4

Training results

Training Loss Epoch Step Validation Loss Precision Recall Macro F1 Micro F1 Accuracy Marker F1 Marker Precision Marker Recall Reference F1 Reference Precision Reference Recall Term F1 Term Precision Term Recall
0.0857 1.0 3125 0.0966 0.8374 0.7889 0.8124 0.8124 0.9676 0.6143 0.5874 0.6437 0.9628 0.9423 0.9842 0.8291 0.8656 0.7955
0.058 2.0 6250 0.0606 0.8869 0.9146 0.9006 0.9006 0.9814 0.9405 0.9126 0.9702 0.9689 0.9511 0.9873 0.8923 0.8805 0.9045
0.0415 3.0 9375 0.0642 0.9077 0.9131 0.9104 0.9104 0.9823 0.9524 0.9262 0.9801 0.9742 0.9614 0.9873 0.9021 0.9026 0.9016
0.0283 4.0 12500 0.0646 0.9066 0.9089 0.9077 0.9077 0.9819 0.9564 0.9326 0.9815 0.9712 0.9555 0.9873 0.8986 0.9008 0.8965

Framework versions

  • Transformers 4.21.3
  • Pytorch 1.12.1+cu113
  • Datasets 2.4.0
  • Tokenizers 0.12.1
Downloads last month
7
Safetensors
Model size
65.2M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.