eole-config.yaml

## IO
save_data: data
overwrite: True
seed: 1234
report_every: 100
valid_metrics: ["BLEU"]
tensorboard: true
tensorboard_log_dir: tensorboard

### Vocab
src_vocab: fren/fr.eole.vocab
tgt_vocab: fren/en.eole.vocab
src_vocab_size: 32000
tgt_vocab_size: 32000
vocab_size_multiple: 8
share_vocab: false
n_sample: 0

data:
    corpus_1:
        path_src: fren/train.cleaned.filtered.fr
        path_tgt: fren/train.cleaned.filtered.en
        weight: 200
    corpus_2:
        path_src: ../data/newscrawl.backtrans.cleaned.filtered.fr
        path_tgt: ../data/newscrawl.backtrans.cleaned.filtered.en
        weight: 35
    corpus_3:
        path_src: ../data/madlad.backtrans.cleaned.filtered.fr
        path_tgt: ../data/madlad.backtrans.cleaned.filtered.en
        weight: 68
    corpus_4:
        path_src: ../data/hansard.fr
        path_tgt: ../data/hansard.en
        weight: 5
    valid:
        path_src: fren/dev.fr
        path_tgt: fren/dev.en

transforms: [sentencepiece, filtertoolong]
transforms_configs:
  sentencepiece:
    src_subword_model: "fren/fr.spm.model"
    tgt_subword_model: "fren/en.spm.model"
  filtertoolong:
    src_seq_length: 256
    tgt_seq_length: 256

training:
    # Run configuration
    model_path: quickmt-fr-en-eole-model
    keep_checkpoint: 4
    train_steps: 200000
    save_checkpoint_steps: 5000
    valid_steps: 5000
    
    # Train on a single GPU
    world_size: 1
    gpu_ranks: [0]

    # Batching 120,000 tokens
    # For RTX 5090, 15000 batch size, accum_count 8
    batch_type: "tokens"
    batch_size: 6000
    valid_batch_size: 2048
    batch_size_multiple: 8
    accum_count: [20]
    accum_steps: [0]

    # Optimizer & Compute
    compute_dtype: "fp16"
    optim: "adamw"
    #use_amp: True
    learning_rate: 3.0
    warmup_steps: 5000
    decay_method: "noam"
    adam_beta2: 0.998

    # Data loading
    bucket_size: 256000
    num_workers: 4
    prefetch_factor: 128
    
    # Hyperparams
    dropout_steps: [0]
    dropout: [0.1]
    attention_dropout: [0.1]
    max_grad_norm: 0
    label_smoothing: 0.1
    average_decay: 0.0001
    param_init_method: xavier_uniform
    normalization: "tokens"

model:
    architecture: "transformer"
    share_embeddings: false
    share_decoder_embeddings: false
    add_estimator: false
    add_ffnbias: true
    add_qkvbias: false
    layer_norm: standard
    mlp_activation_fn: gelu
    hidden_size: 768
    encoder:
        layers: 12
    decoder:
        layers: 2
    heads: 16
    transformer_ff: 4096
    embeddings:
        word_vec_size: 768
        position_encoding_type: "SinusoidalInterleaved"