# 🤖 PIPELINE DEFINITION
nlp:
  "@core": pipeline

  lang: fr

  components:
    parser:
      '@factory': eds.biaffine_dep_parser
      hidden_size: 64
      decoding_mode: greedy
      dropout_p: 0.
      use_attrs: ['pos_']

      embedding:
        '@factory': eds.transformer
        model: hf-internal-testing/tiny-bert
        window: 512
        stride: 256

# 📈 SCORERS
scorer:
  speed: false
  dep:
    '@metrics': "eds.dep_parsing"

# 🎛️ OPTIMIZER
optimizer:
  optim: adamw
  module: ${ nlp }
  total_steps: ${ train.max_steps }
  groups:
    ".*":
      lr: 1e-3

# 📚 DATA
train_data:
  data:
    "@readers": conll
    path: ./rhapsodie_sample.conllu
  shuffle: dataset
  batch_size: 1 docs
  pipe_names: [ "parser" ]

val_data:
  "@readers": conll
  path: ./rhapsodie_sample.conllu

# 🚀 TRAIN SCRIPT OPTIONS
train:
  nlp: ${ nlp }
  train_data: ${ train_data }
  val_data: ${ val_data }
  max_steps: 20
  validation_interval: 10
  max_grad_norm: 5.0
  scorer: ${ scorer }
  num_workers: 0
  optimizer: ${ optimizer }
  grad_dev_policy: "clip_mean"
  log_weight_grads: true