mix nasty.fine_tune.pos (Nasty v0.3.0)

Fine-tunes a pre-trained transformer model for POS tagging.

Usage

mix nasty.fine_tune.pos \
  --model roberta_base \
  --train data/en_ewt-ud-train.conllu \
  --validation data/en_ewt-ud-dev.conllu \
  --output models/pos_finetuned \
  --epochs 3 \
  --batch-size 16

Options

• --model - Base transformer model to fine-tune (required) Options: bert_base_cased, roberta_base, xlm_roberta_base
• --train - Path to training data in CoNLL-U format (required)
• --validation - Path to validation data (optional)
• --output - Output directory for fine-tuned model (default: priv/models/finetuned)
• --epochs - Number of training epochs (default: 3)
• --batch-size - Training batch size (default: 16)
• --learning-rate - Learning rate (default: 3e-5)
• --max-length - Maximum sequence length (default: 512)
• --eval-steps - Evaluate every N steps (default: 500)

Examples

# Quick fine-tuning with defaults
mix nasty.fine_tune.pos --model roberta_base --train data/train.conllu

# Full configuration
mix nasty.fine_tune.pos \
  --model bert_base_cased \
  --train data/train.conllu \
  --validation data/dev.conllu \
  --epochs 5 \
  --batch-size 32 \
  --learning-rate 0.00002 \
  --output models/pos_bert_finetuned