Task_type parameter of LoraConfig

🤗Transformers
1

I am training a fine-tune of codellama using PEFT but not sure how to use the task_type parameter of LoraConfig. Should it be CAUSAL_LM or SEQ_2_SEQ_LM or something else? Does it have any affect?

The goal of my model is to parse an input for independent clauses in a sentence. For example, it would insert a delimiter, such as in this sentence: “the tea was on the stove and was at high temperature” , separating the independent clause from the subordinate clause. My training data is all in a single col and each row looks like this (where the → and are custom tokens I add to the tokenizer vocab and the is the EOS token):

“the tea was on the stove and was at high temperature → the tea was on the stove and was at high temperature ”

2 Likes
2

Does task_type matters in LoraConfig, and if so, in what way?

1 Like
3

I have the same question. Does task_type matter?

2 Likes
4

In the source code, task_type is not even there:

1 Like
5

The task_type parameter is used in the superclass PeftConfig

5 Likes
6 
    SEQ_CLS = "SEQ_CLS"
    SEQ_2_SEQ_LM = "SEQ_2_SEQ_LM"
    CAUSAL_LM = "CAUSAL_LM"
    TOKEN_CLS = "TOKEN_CLS"
    QUESTION_ANS = "QUESTION_ANS"
    FEATURE_EXTRACTION = "FEATURE_EXTRACTION"

    Overview of the supported task types:
    - SEQ_CLS: Text classification.
    - SEQ_2_SEQ_LM: Sequence-to-sequence language modeling.
    - Causal LM: Causal language modeling.
    - TOKEN_CLS: Token classification.
    - QUESTION_ANS: Question answering.
    - FEATURE_EXTRACTION: Feature extraction. Provides the hidden states which can be used as embeddings or features
      for downstream tasks.

See here for source: peft/src/peft/utils/peft_types.py at v0.8.2 · huggingface/peft · GitHub

7 Likes
7

Late reply but I was wondering the same and could not find details in the doc so I put an answer here.

Based on Peft git repo, it seems that the model class returned by get_peft_model(peft_config)
depends on this value :
peft/src/peft/mapping , with the following mapping Line 81 :

MODEL_TYPE_TO_PEFT_MODEL_MAPPING: dict[str, type[PeftModel]] = {
    "SEQ_CLS": PeftModelForSequenceClassification,
    "SEQ_2_SEQ_LM": PeftModelForSeq2SeqLM,
    "CAUSAL_LM": PeftModelForCausalLM,
    "TOKEN_CLS": PeftModelForTokenClassification,
    "QUESTION_ANS": PeftModelForQuestionAnswering,
    "FEATURE_EXTRACTION": PeftModelForFeatureExtraction,
}

If you do not specify a value, you get a default PeftModel (Lines 210-220)

    if peft_config.task_type not in MODEL_TYPE_TO_PEFT_MODEL_MAPPING.keys() and not peft_config.is_prompt_learning:
        return PeftModel(
            model,
            peft_config,
            adapter_name=adapter_name,
            autocast_adapter_dtype=autocast_adapter_dtype,
            low_cpu_mem_usage=low_cpu_mem_usage,
        )

    if peft_config.is_prompt_learning:
        peft_config = _prepare_prompt_learning_config(peft_config, model_config)
    return MODEL_TYPE_TO_PEFT_MODEL_MAPPING[peft_config.task_type](
        model, peft_config, adapter_name=adapter_name, autocast_adapter_dtype=autocast_adapter_dtype
    )

Finally it seems that in the case where task_type=SEQ_CLS , the classifier heads are excluded from LoRA : in peft/src/peft/tuners/tuner_utils:

    if output_emb is not None:
        # ignore the last classification head for text generation models
        last_module_name = [name for name, module in model.named_modules() if module is output_emb][0]
        module_names_to_exclude.add(last_module_name)
    elif peft_config.task_type == TaskType.SEQ_CLS:
        # ignore classifier head for classification models (issue 2027)
        # there is no fix name for the classifier head, so check the common ones
        for name in SEQ_CLS_HEAD_NAMES:
            cls_head = getattr(model, name, None)
            if cls_head is not None:
                last_module_name = [name for name, module in model.named_modules() if module is cls_head][0]
                module_names_to_exclude.add(last_module_name)
                break

and also the number of transformer_submodules increases from 1 to 2 in the case of SEQ_2_SEQ_LM (in peft/src/peft/peft_model.py):

        if config.num_transformer_submodules is None:
            config.num_transformer_submodules = 2 if config.task_type == TaskType.SEQ_2_SEQ_LM else 1
2 Likes