Group test covering tokenization strategies, TF-IDF mechanics, stemming vs lemmatization, distributional semantics, Word2Vec/GloVe, and FastText. Tests whether you can reason about representation tradeoffs — not just recall definitions.

Group test covering POS tagging, dependency parsing, co-reference resolution, n-gram models, perplexity, and smoothing. Tests your ability to reason about why pipeline systems fail and what the Markov assumption can and cannot model.

Group test on RNN vanishing gradients, LSTM gating, GRU tradeoffs, BiLSTM representations, teacher forcing, Bahdanau vs Luong attention, and coverage problems. Tests your model of how sequence encoders fail and what attention actually solves.

Group test on BERT pretraining objectives, segment embeddings, RoBERTa/DistilBERT/ALBERT design choices, fine-tuning behavior, and classification system design — including class imbalance, multi-label heads, calibration, and zero-shot strategies.

Group test on BIO tagging, CRF vs linear heads, entity-level evaluation, span-based NER, SQuAD EM/F1 metrics, BERT QA mechanics, SQuAD 2.0 unanswerability, and RAG-based systems. Tests whether you understand why token accuracy masks entity-level failures.

Group test on NMT architecture bottlenecks, attention alignment patterns, BLEU limitations, back-translation, hallucination causes, greedy vs beam vs sampling decoding, temperature, top-k/top-p tradeoffs, and repetition. Tests your ability to pick the right decoding strategy for the right application.

Simulates an easy NLP screening interview. Covers tokenization, word embeddings, classical tasks, language models, sequence models, and attention — all framed as engineering decisions. Includes common reasoning traps that catch underprepared candidates.

Second easy NLP screening simulation. Covers BERT intuition, classification, NER basics, QA retrieval, MT evaluation, and text generation — all through practical scenario framing. A distinct question set from Mock I, no overlap.

Simulates a medium-difficulty NLP engineering interview. Tests feature space tradeoffs, GloVe vs Word2Vec performance gaps, perplexity domain shifts, attention complexity, zero-shot NLI transfer, production class-deletion risks, and multi-hop QA architecture. Requires multi-step reasoning throughout.

Second medium NLP interview simulation — distinct question set. Covers TF-IDF semantic gaps, CRF vs per-token softmax, BiLSTM pooling for long docs, NLI zero-shot, cross-lingual NER transfer, back-translation, and top-p diversity tradeoffs. FAANG-style scenario framing throughout.

Simulates a hard NLP engineering interview targeting senior/staff candidates. Questions involve domain-adaptive pretraining, semantic drift over time, pipeline error propagation, variational dropout, monotonic attention failure modes, BERT layer probing, concept vs data drift, NER production tradeoffs, and dense retrieval gaps. Expect traps.

Second hard NLP engineering simulation — no question overlap with Hard Mock I. Covers BPE morphological fragmentation, LM length bias in ASR, Arabic morphology tokenization, constrained decoding complexity, TF-IDF/BPE interactions, hard CRF boundary questions, extractive QA multi-hop failure, and adversarial MT evaluation. All questions require multi-step reasoning.

Senior ML engineer / AI architect-level assessment. 18 questions drawn from the hardest material across all 12 NLP topics. Tests deceptive edge cases, production failure modes, and multi-concept intersections: domain pretraining vs tokenization, LM length bias in ASR, BPE-morphology interactions, BERT probing hierarchy, pipeline error propagation, precision-recall production tradeoffs, and constrained decoding complexity. No easy warmup — this starts hard and stays hard.

Staff engineer / advanced ML interview gauntlet. 18 questions spanning every NLP subsystem — all hard, all distinct from Elite Test I. Covers recurrent dropout temporal inconsistency, monotonic attention language-pair failure, BERT layer hierarchy implications, semantic drift over time, variational dropout internals, multi-hop QA architectural limits, Arabic NMT morphology, CRF pipeline error propagation, coverage and hallucination in MT, and statistical LM OOV edge cases. Requires holding multiple system-level concepts simultaneously.