lava-policy-multiwoz

This is the best performing LAVA_kl model from the LAVA paper which can be used as a word-level policy module in ConvLab3 pipeline.

Refer to ConvLab-3 for model description and usage.

Training procedure

The model was trained on MultiWOZ 2.0 data using the LAVA codebase. The model started with VAE pre-training and fine-tuning with informative prior KL loss, followed by corpus-based RL with REINFORCE.

Training hyperparameters

The following hyperparameters were used during SL training:

• y_size: 10
• k_size: 20
• beta: 0.1
• simple_posterior: true
• contextual_posterior: false
• learning_rate: 1e-03
• max_vocab_size: 1000
• max_utt_len: 50
• max_dec_len: 30
• backward_size: 2
• train_batch_size: 128
• seed: 58
• optimizer: Adam
• num_epoch: 100 with early stopping based on validation set

The following hyperparameters were used during RL training:

• tune_pi_only: false
• max_words: 100
• temperature: 1.0
• episode_repeat: 1.0
• rl_lr: 0.01
• momentum: 0.0
• nesterov: false
• gamma: 0.99
• rl_clip: 5.0
• random_seed: 38