Item analysis
Question statistics
| Question | n | Difficulty p | Discrimination r_pb | Verdict |
|---|---|---|---|---|
| What is the main role of the softmax operation in the Bahdanau attention mechanism? r_pb=-0.15 < 0.15 — does not separate strong from weak | 6 | 0.83 | -0.15 | low discrimination |
| What is the main advantage of using Byte-Pair Encoding (BPE) for tokenization in Transformers? | 6 | 0.50 | 0.78 | OK |
| Why does the LSTM's cell state update help prevent vanishing gradients for long-term dependencies? | 6 | 0.67 | 0.75 | OK |
| Why does the Transformer decoder use a causal mask during training? | 6 | 0.83 | 0.53 | OK |
| Why is it essential to scale the dot products by 1/√d in the Transformer's self-attention? | 6 | 0.50 | 0.41 | OK |
Rejection rule (from 5 attempts): a question is flagged when its difficulty p falls outside 0.30–0.85 (share answering correctly) or its point-biserial discrimination drops below 0.15 (correlation with the rest of the attempt score). Retiring and regenerating flagged questions runs offline: uv run python -m app.quality … --regenerate.