Model hosting

[JViggiani]

Hello! Very interesting implementation. Do you have a trained model hosted available anywhere for downloading?

Thank you

[bosung]

Hi, thank you for your interest in ZETT! Currently, we don't have any published models online, but training a model takes less than 10 minutes for each split (based on TITAN RTX 24GB). Hope this helps!

[JViggiani]

That's okay, I managed to get it working. I'm very impressed by the training time and the speed at which it trained.

I have a somewhat unrelated question around the automatic template generation. I see in wrapper.py there is a function:

    def load_test_auto_templ(self):
        raise NotImplementedError("Use this class: AutoTemplExtractor")

And this is called in:

    def generate_entity_span(self, 
                            data: Sentence, 
                            generator: TextGenerator, 
                            tokenizer: Union[PreTrainedTokenizerFast, AutoTokenizer], 
                            target_labels: List[str], 
                            mode: str, 
                            task_type: str, 
                            use_label_constraint: bool) -> List[dict]:
      # ...

        if mode in ["auto_templ_single", "auto_templ_multi"]:
            test_auto_templs = self.load_test_auto_templ()

    # ...

                for k in final_target_labels:
                if mode in ["auto_templ_single", "auto_templ_multi"]:
                    input_text = []
                    for nt in range(min(self.top_n_templ, len(test_auto_templs[k]))):
                        _templ = test_auto_templs[k][nt]
                        # ...
                else:
                    _templ = self.relname2template[k]   
                    # ... 

Is this automatic template generation implemented anywhere and hosted somewhere else? Thank you! :)

[bosung]

This is the code for auto-generated templates by LLM, which was one of our experiments, but not included to the paper since the performance was not good enough. Maybe I should've deleted this part -- sorry for your confusion!