What are the bare-bone template of the Bonito?

[pacozaa]

I would like to use this model with Ollama or llama.cpp but I would like to know the bare-bone explanation of bonito's template. Would you mind giving a short explanation?

[pacozaa]

Ah ha, your paper explain it!

<|tasktype|>
Yes-no question answering
<|context|>
Zinedine Zidane -- After retiring as a player, Zidane
transitioned into coaching, becoming assistant coach at
Real Madrid… after the victory, he resigned as Real
Madrid coach.
<|task|>

Still don't mind more explanation and examples though

Cheers!

[Nihal Nayak]

You are right. We have included the template in the paper. We have also included the preprocessing step in abstract.py. Please look at the following lines of code.

bonito/bonito/abstract.py

Lines 26 to 71 in 0b6b23d

     def _prepare_bonito_input( 
         self, context_dataset: Dataset, task_type: str, context_col: str, **kwargs 
     ) -> Dataset: 
         """ 
         Prepares the input for the Bonito model. 
  
         This method takes a context dataset, a task type, and a context 
         column name, and prepares the dataset for the Bonito model. 
         If the task type is not recognized, it raises a ValueError. 
  
         Args: 
             context_dataset (Dataset): The dataset that provides the 
                 context for the task. 
             task_type (str): The type of the task. This can be a 
                 short form or a full form. If the task type is not 
                 recognized, a ValueError is raised. 
             context_col (str): The name of the column in the dataset 
                 that provides the context for the task. 
             **kwargs: Additional keyword arguments. 
  
         Returns: 
             Dataset: The prepared dataset for the Bonito model. 
         """ 
         # get the task type name 
         if task_type in SHORTFORM_TO_FULL_TASK_TYPES.values(): 
             full_task_type = task_type 
         elif task_type in SHORTFORM_TO_FULL_TASK_TYPES: 
             full_task_type = SHORTFORM_TO_FULL_TASK_TYPES[task_type] 
         else: 
             raise ValueError(f"Task type {task_type} not recognized") 
  
         def process(example): 
             input_text = "<|tasktype|>\n" + full_task_type.strip() 
             input_text += ( 
                 "\n<|context|>\n" + example[context_col].strip() + "\n<|task|>\n" 
             ) 
             return { 
                 "input": input_text, 
             } 
  
         return context_dataset.map( 
             process, 
             remove_columns=context_dataset.column_names, 
             num_proc=kwargs.get("num_proc", 1), 
         ) 
  

Hope this helps! 😄

[pacozaa]

In case anyone stumbles on this here is Ollama library you can run https://ollama.com/pacozaa/bonito

And quantize and convert to gguf article here
https://medium.com/@sarinsuriyakoon/convert-pytorch-model-to-quantize-gguf-to-run-on-ollama-5c5dbc458208