NCP-GENL Fine-Tuning Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  The checkpoint was saved and later loaded with a mismatched tokenizer vocabulary, so input token IDs no longer map to the trained embeddings.

  Why it's wrong here

  A tokenizer mismatch would typically cause obvious garbage or errors at inference, not coherent-but-repetitive text. It also would not be consistent with a smoothly decreasing training loss on the same data pipeline. This is a plausible but incorrect cause for the observed repetition behavior.

• ✗

  The evaluation prompts were longer than the maximum sequence length used during fine-tuning, so the model never learned to handle them.

  Why it's wrong here

  Length mismatch can degrade quality on unusually long prompts, but the reported symptom is repetition across outputs, which is broader than a length generalization issue. Fine-tuning on shorter sequences does not typically cause pervasive token repetition on normal-length prompts. This explanation does not account for the coherent-but-repetitive pattern.

• ✓

  The model was trained with a sequence packing strategy that concatenated unrelated samples without proper attention masking between them.

  Why this is correct

  Sequence packing can improve throughput, but if attention masking does not prevent tokens from attending across sample boundaries, the model learns spurious cross-sample dependencies. This degrades generation quality, often producing repetition and incoherence, even though the training loss looks healthy because the model is fitting the corrupted attention pattern.

• ✗

  The learning rate was too high, causing the optimizer to overshoot and permanently corrupt the base model weights.

  Why it's wrong here

  An excessively high learning rate usually produces loss spikes or divergence, which the engineer would have observed. A smoothly decreasing loss is inconsistent with that failure mode. Overshooting also tends to produce instability during training rather than clean completion followed by repetitive inference output.