apply_chat_template_config.yaml

settings:
  src_path: data/smol-smoltalk_train_first_10K.jsonl
  dst_path: prepared_data/smol-smoltalk_train_first_10K.jsonl
  messages_key: messages
  pbin_creation_config_file_path: configs/packed_chat_dataset_config.yaml
  split_config:
    splitting:
      train: 70
      val: 15
      test: 15
    seed: 1234

instruction_data_transformation:
  role_mapping:
    system: System
    user: User
    assistant: Assistant

# The b_include_to_loss_token, e_include_to_loss_token are required to be part of each chat template for proper loss masking!
# messages is a required special "chat_template_data" and corresponds to the data column settings.messages_key
jinja2_chat_template: |
    {{ chat_template_data.system_instruction + '\n' }}
    {% for turn in messages %}
    {{ turn.role + ':' }}
    {% if turn.role == chat_template_data.assistant_role %}
    {{ chat_template_data.special_tokens.b_include_to_loss_token}}
    {% else %}
    {{ " " }}
    {% endif %}
    {{ turn.content + '\n'}}
    {% if turn.role == chat_template_data.assistant_role %}
    {{ chat_template_data.special_tokens.e_assistant_token}}
    {{ chat_template_data.special_tokens.e_include_to_loss_token}}
    {% endif %}
    {% endfor %}

# The key-value pairs of chat_template_data are passed to the Jinja2 template and 
# are not type checked for full compliance with the chat tempalate!
chat_template_data:
  assistant_role: Assistant
  system_instruction: "You are Mody, a helpful assistant trained by the modalities team. Answer friendly and informatively to the user's messages."
  # Currently only works with HF tokenizers, as the special tokens are added to the tokenizer
  special_tokens:
      b_include_to_loss_token: <|im_start|>
      e_include_to_loss_token: <|im_end|>
      e_assistant_token: <|endoftext|>