Missing component when run

Beginners
1

Hello, I’m new to the topic and currently following the LLM Course. But I got this error saying the component is missing even though it’s a snippet from the course

image
image2048×1088 315 KB

1 Like
2

that’s maybe harmless warning message.

What this message means in your case (it’s not a “missing code component”)

From the screenshot, you are doing the LLM Course step where you load:

  • checkpoint = "bert-base-uncased"
  • AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)

Transformers then prints a weight loading report:

  • UNEXPECTED weights like cls.predictions.* and cls.seq_relationship.*
  • MISSING weights like classifier.weight and classifier.bias

This is a normal warning/report when you load a base pretrained BERT checkpoint into a task-specific model class (sequence classification). It does not mean a Python module or layer is “missing” from your environment. The LLM Course explicitly says you should see this warning at that point, and explains why. (Hugging Face)

Background (why these specific keys appear)

1) BERT pretraining vs downstream tasks

The checkpoint bert-base-uncased is pretrained mainly for Masked Language Modeling (MLM) and Next Sentence Prediction (NSP). Those correspond to weights under:

  • cls.predictions.* (MLM head)
  • cls.seq_relationship.* (NSP head)

When you load AutoModelForSequenceClassification, you are asking Transformers to build a model whose “head” is a classifier, not MLM/NSP. So those MLM/NSP head weights are present in the checkpoint but not used by your chosen architecture → reported as UNEXPECTED. This exact “warning is expected when switching architecture/task heads” is described in the canonical Transformers issue and in HF discussions. (GitHub)

2) Why classifier.* is “missing”

BertForSequenceClassification (and AutoModelForSequenceClassification) includes a new final layer:

  • classifier.weight
  • classifier.bias

Those do not exist in the base checkpoint because the base checkpoint is not trained for your classification labels. So Transformers creates them randomly and reports them as MISSING (newly initialized). HF maintainers explicitly state this is normal and that you must fine-tune before using it for inference. (Hugging Face Forums)

Why the LLM Course shows this exact output

The LLM Course chapter you’re following says (paraphrased):

  • you’ll get warnings now (unlike earlier chapters)
  • the pretrained head is discarded
  • a new sequence classification head is added
  • some weights are unused, and the new head is randomly initialized
  • you should proceed to training

That description matches your screenshot precisely. (Hugging Face)

What you should do next (choose based on your goal)

A) If you’re following the course and you will train

Do nothing special. Continue to Trainer(...).train().

This warning is basically saying: “the classifier head starts from scratch; training will learn it.” (Hugging Face)

B) If you expected to run inference immediately (no training)

Then this warning is a problem for your goal, because the classifier head is random.

Fix: Load a model checkpoint that is already fine-tuned for sequence classification (task-specific checkpoint), not bert-base-uncased. The warning exists specifically to prevent people from assuming the model is ready for predictions. (Hugging Face Forums)

C) If you already trained but still see “MISSING classifier.*”

Most common cause: you are still loading bert-base-uncased again instead of loading the directory/repo that contains your fine-tuned checkpoint.

Fix: load from your saved model output directory (the one created by the Trainer / save_pretrained()), not from the original base checkpoint. The GitHub issue calls this out as the “NOT expected” scenario when you believe you’re loading an identical architecture/checkpoint. (GitHub)

Quick sanity check (to confirm it’s “expected” and nothing else is wrong)

Use output_loading_info=True to see exactly what’s missing/unexpected. The docs describe this parameter as returning missing/unexpected keys. (Hugging Face)

Expected for your case:

  • missing_keys contains only classifier.weight and classifier.bias
  • unexpected_keys contains only the pretraining head keys under cls.*

If you see large parts of bert.encoder.* or embeddings missing, that would indicate a real mismatch (wrong model family/checkpoint).

Optional: suppress the warning (only after you understand it)

If the output is noisy, you can lower Transformers logging verbosity. (This doesn’t change behavior, only printing.) (Hugging Face Forums)

Bottom line for your screenshot

This is the expected warning for the LLM Course sequence classification fine-tuning step: the classification head is new and must be trained, and the MLM/NSP heads from the base checkpoint are not used. (Hugging Face)