Tips & FAQ

# Common mistakes on the Hugging Face NLP Certification

common_mistakes = [
    {
        "mistake": "Using AutoModel instead of AutoModelForSequenceClassification",
        "impact": "No classification head - model outputs hidden states, not logits",
        "fix": "Always use the task-specific AutoModel variant"
    },
    {
        "mistake": "Forgetting to set return_tensors in tokenizer",
        "impact": "Returns Python lists instead of tensors - model cannot process",
        "fix": "Always pass return_tensors='pt' (PyTorch) or 'tf' (TensorFlow)"
    },
    {
        "mistake": "Not aligning labels with subword tokens for NER",
        "impact": "Label mismatch causes training errors or poor performance",
        "fix": "Use word_ids() to map labels, set -100 for special/sub tokens"
    },
    {
        "mistake": "Confusing evaluation_strategy with save_strategy",
        "impact": "Model may not save the best checkpoint",
        "fix": "Set both to the same value ('epoch' or 'steps')"
    },
    {
        "mistake": "Not calling tokenizer.push_to_hub() alongside model",
        "impact": "Users cannot load your model because tokenizer files are missing",
        "fix": "Always push both model and tokenizer"
    },
    {
        "mistake": "Using wrong model architecture for the task",
        "impact": "Poor performance or errors",
        "fix": "Encoder for understanding, decoder for generation, seq2seq for translation"
    },
    {
        "mistake": "Not setting num_labels when loading classification model",
        "impact": "Default num_labels=2, causes error for multi-class datasets",
        "fix": "Always pass num_labels matching your dataset"
    },
    {
        "mistake": "Forgetting padding=True for batched inference",
        "impact": "Sequences of different lengths cannot be batched",
        "fix": "Always use padding=True when processing multiple examples"
    }
]

# Quick reference - review before the exam

quick_ref = {
    "Classification": {
        "pipeline": "text-classification",
        "model": "AutoModelForSequenceClassification",
        "metric": "accuracy, f1"
    },
    "NER": {
        "pipeline": "ner (or token-classification)",
        "model": "AutoModelForTokenClassification",
        "metric": "seqeval (precision, recall, F1)"
    },
    "QA": {
        "pipeline": "question-answering",
        "model": "AutoModelForQuestionAnswering",
        "metric": "squad (exact_match, F1)"
    },
    "Summarization": {
        "pipeline": "summarization",
        "model": "AutoModelForSeq2SeqLM",
        "metric": "rouge (rouge1, rouge2, rougeL)"
    },
    "Translation": {
        "pipeline": "translation",
        "model": "AutoModelForSeq2SeqLM",
        "metric": "bleu, sacrebleu"
    },
    "Generation": {
        "pipeline": "text-generation",
        "model": "AutoModelForCausalLM",
        "metric": "perplexity"
    }
}