{
  "lesson": "62-vision-language-pretraining",
  "title": "Vision-Language Pretraining",
  "questions": [
    {
      "stage": "pre",
      "question": "What two distinct capabilities does combining InfoNCE and LM loss train into one model?",
      "options": [
        "Ranking (find the right image for a caption) and generation (write a caption for an image)",
        "Compression and decompression",
        "Speed and accuracy",
        "Tokenization and detokenization"
      ],
      "correct": 0,
      "explanation": "InfoNCE handles ranking; LM loss handles generation. CoCa, BLIP, and friends combine both in one training pass."
    },
    {
      "stage": "pre",
      "question": "In InfoNCE for a batch of N pairs, how many positive pairs and negative pairs are formed?",
      "options": [
        "N positives along the diagonal of the similarity matrix and N*(N-1) negatives off-diagonal",
        "N positives and N negatives",
        "1 positive and N negatives",
        "N^2 positives"
      ],
      "correct": 0,
      "explanation": "The N matching pairs sit on the diagonal; everything off-diagonal is a negative."
    },
    {
      "stage": "check",
      "question": "Why is the contrastive temperature usually a learned parameter?",
      "options": [
        "It avoids using Adam",
        "PyTorch requires it",
        "Learned tau is faster on GPU",
        "Hand-tuning is fragile; letting tau learn allows the model to find the right softmax peakedness as the embedding scale evolves during training"
      ],
      "correct": 3,
      "explanation": "CLIP introduced learned log-tau so the contrastive softmax stays in the useful regime as embedding norms grow."
    },
    {
      "stage": "check",
      "question": "Which sub-modules receive gradient from the LM loss but not from the contrastive loss?",
      "options": [
        "The MLP projector",
        "The vision encoder",
        "The text-side encoder",
        "The cross-attention decoder"
      ],
      "correct": 3,
      "explanation": "Contrastive loss does not touch the decoder; LM loss does not touch the text-side mean-pool encoder."
    },
    {
      "stage": "check",
      "question": "What does ignore_index=PAD_ID accomplish in the LM cross-entropy?",
      "options": [
        "Speeds up the loss",
        "Encrypts pad tokens",
        "Masks padding positions out of the loss so the model is not punished for predicting tokens after the caption ends",
        "Reorders tokens"
      ],
      "correct": 2,
      "explanation": "Padding positions carry no signal; ignoring them stops the loss from learning to predict pad."
    },
    {
      "stage": "post",
      "question": "Why is a 50-step demo on synthetic data enough to call the lesson complete?",
      "options": [
        "The model reaches state of the art at 50 steps",
        "50 steps is the maximum supported",
        "The demo proves the gradient plumbing works end to end; production training shape is identical but runs for millions of steps on real data",
        "Real data would change the architecture"
      ],
      "correct": 2,
      "explanation": "The dynamics demonstrated (both losses decrease, no NaN, stable tau) are the same dynamics a billion-image run would show; just scaled up."
    }
  ]
}