Pandamtl

Add lightweight task heads (linear layers) on top of the shared encoder’s outputs. Example for POS tagging:

class PandaMTLModel(PreTrainedModel):
    def __init__(self, base_transformer, num_pos_labels):
        super().__init__()
        self.transformer = base_transformer
        self.pos_head = nn.Linear(hidden_size, num_pos_labels)
def forward(self, input_ids, labels_translation=None, labels_pos=None):
    encoder_outputs = self.transformer.encoder(input_ids)
    translation_logits = self.transformer.decoder(encoder_outputs)
    pos_logits = self.pos_head(encoder_outputs.last_hidden_state)
    # Combine losses...

Given Montreal’s reputation for art and design (Mural Festival, Just for Laughs), pandamtl could be a tiny branding studio or a Shopify storefront. The "panda" suggests a friendly, black-and-white minimalist aesthetic. Searchers might be looking for:

# Pseudocode using a Hugging Face-like interface
from pandamtl import PandaMTLModel, PandaMTLConfig
config = PandaMTLConfig(
tasks=["translation", "pos", "ner"],
task_weights=[0.7, 0.2, 0.1],
shared_encoder_layers=6,
decoder_layers=6
) pandamtl
model = PandaMTLModel.from_pretrained("pandamtl-base-en-fr")
train_dataset = load_mtl_dataset("en-fr", tasks=["translation", "pos", "ner"])
trainer = PandaMTLTrainer(model, train_dataset, learning_rate=3e-5)
trainer.train()
 Add lightweight task heads (linear layers) on top