{
  "module": "FTDD-08 — Qwen3",
  "course": "3 — LLM Fine-Tuning Masterclass",
  "version": "1.0.0",
  "duration_minutes": 45,
  "total_questions": 10,
  "bloom_distribution": {
    "target": "40% recall / 30% application / 30% analysis",
    "actual": { "recall": 4, "application": 3, "analysis": 3 }
  },
  "passing_score_percent": 70,
  "questions": [
    {
      "id": "Q01", "bloom": "recall", "type": "multiple_choice",
      "prompt": "What is thinking-mode fusion (Stage 3 of the Qwen3 pipeline)?",
      "options": [
        "A router that decides whether to send a query to a thinking model or a non-thinking model.",
        "Merging a thinking (long-CoT) model and a non-thinking (fast) model into a single set of weights, with a control signal selecting which mode the model expresses.",
        "Training two separate models and load-balancing between them at serving time.",
        "Distilling a large reasoning model into a smaller one via SFT."
      ],
      "answer_index": 1,
      "rationale": "Fusion is the novel Stage 3: one set of weights with both modes coexisting, selected by a control token or budget. The key insight is that thinking and direct-response pathways are not mutually exclusive in the weights. This is what R1 does NOT do — R1 ships a pure reasoner."
    },
    {
      "id": "Q02", "bloom": "recall", "type": "multiple_choice",
      "prompt": "Name the four stages of the Qwen3 post-training pipeline in order.",
      "options": [
        "Pretraining → SFT → DPO → GRPO",
        "Long-CoT cold start → reasoning RL → thinking-mode fusion → general RL",
        "Thinking-mode fusion → long-CoT cold start → general RL → reasoning RL",
        "Reasoning RL → general RL → thinking-mode fusion → cold start"
      ],
      "answer_index": 1,
      "rationale": "(1) Long-CoT cold start seeds reasoning format. (2) Reasoning RL (GRPO-style) scales capability. (3) Thinking-mode fusion merges both modes. (4) General RL aligns the full distribution. Stages 1 and 2 echo R1; Stage 3 is the novel contribution."
    },
    {
      "id": "Q03", "bloom": "recall", "type": "multiple_choice",
      "prompt": "What does the thinking budget control?",
      "options": [
        "The number of parameters active per token in the MoE.",
        "How many tokens the model may spend in its <think> deliberation block before answering.",
        "The learning rate during RL training.",
        "The number of GPUs allocated to serving."
      ],
      "answer_index": 1,
      "rationale": "The thinking budget is an inference-time parameter controlling how many tokens the model may spend in <think> deliberation. Low = short/empty think, fast direct answer. High = long self-correcting chain. Same weights serve both modes."
    },
    {
      "id": "Q04", "bloom": "recall", "type": "multiple_choice",
      "prompt": "What is Qwen3's pretraining scale, and what model sizes does the family include?",
      "options": [
        "3T tokens; dense models only, 7B and 13B.",
        "36T+ tokens; dense models from 0.6B to 32B and a 235B-A22B MoE flagship.",
        "500B tokens; a single 70B dense model.",
        "15T tokens; dense models from 8B to 405B."
      ],
      "answer_index": 1,
      "rationale": "Qwen3 was pretrained on 36T+ tokens. The family spans dense models (0.6B, 1.7B, 4B, 8B, 14B, 32B) and a Mixture-of-Experts flagship at 235B total / 22B active (A22B). This scale ensures the non-thinking mode inherits genuine capability."
    },
    {
      "id": "Q05", "bloom": "application", "type": "multiple_choice",
      "prompt": "You deployed Qwen3 as a general-purpose chatbot. Users complain it is slow even on simple questions like 'what time is it.' What is most likely wrong?",
      "options": [
        "The model is too large; switch to a smaller dense variant.",
        "The thinking budget is set to max, so every query runs a long thinking chain. Lower the default budget for this workload.",
        "The fusion failed; you need to retrain Stage 3.",
        "The model is broken; redeploy with a different quantization."
      ],
      "answer_index": 1,
      "rationale": "A max thinking budget imposes deliberation on every query, including trivial ones — pure latency tax with no quality gain. The fix is operational, not architectural: set a low default budget for the chatbot workload. The model will answer simple queries directly. Raise per query only when deliberation is warranted."
    },
    {
      "id": "Q06", "bloom": "application", "type": "multiple_choice",
      "prompt": "A batch math-solver processes competition problems overnight. What thinking budget should it use, and why?",
      "options": [
        "Low — batch jobs should minimize token spend.",
        "Zero — math should be solved without deliberation.",
        "High — competition math rewards long self-correcting chains; latency is not critical in batch, accuracy is.",
        "It does not matter — the budget only affects chat, not math."
      ],
      "answer_index": 2,
      "rationale": "Competition math rewards deliberation: self-correction, verification, multi-step reasoning. A high budget gives the model the token headroom to reason. Latency is not critical in an overnight batch job; accuracy is. Set the default high."
    },
    {
      "id": "Q07", "bloom": "application", "type": "multiple_choice",
      "prompt": "Your team is evaluating a Qwen3 deployment and only ran benchmarks in thinking mode. What risk does this create?",
      "options": [
        "No risk — thinking mode is the only mode that matters.",
        "You may miss degradation in non-thinking mode and misjudge the model for production, where most traffic may use the fast mode. Stage 4 (general RL) exists to keep non-thinking strong — you must verify it did.",
        "You will overestimate latency.",
        "You need to retrain the model before evaluation."
      ],
      "answer_index": 1,
      "rationale": "A hybrid model must be evaluated in BOTH modes against their respective benchmarks. Stage 4 exists precisely to keep the non-thinking mode strong. If you evaluate only thinking mode, you will not catch non-thinking degradation and will misjudge the model for the production workload that relies on the fast mode."
    },
    {
      "id": "Q08", "bloom": "analysis", "type": "multiple_choice",
      "prompt": "Why does the thinking budget eliminate the need for a separate query router in a hybrid deployment?",
      "options": [
        "Because the budget automatically classifies queries as hard or easy.",
        "Because it collapses routing into a single continuous parameter: set a default budget for your workload and the model spends within it. You do not need a classifier to decide 'is this query hard' — the budget handles the trade-off directly.",
        "Because Qwen3 cannot produce a wrong routing decision.",
        "Because the budget disables the non-thinking mode entirely."
      ],
      "answer_index": 1,
      "rationale": "Without a budget, a hybrid still needs a per-query router to decide thinking vs non-thinking, and the router is itself a failure mode. The budget collapses routing into one continuous parameter. Set a default for your workload; the model adapts within it. No classifier, no misroutes."
    },
    {
      "id": "Q09", "bloom": "analysis", "type": "multiple_choice",
      "prompt": "Qwen3's Stages 1 and 2 mirror R1's cold-start SFT and reasoning RL. What does the DIVERGENCE at Stage 3 (fusion) tell you about Qwen3's design goal versus R1's?",
      "options": [
        "Qwen3 abandoned reasoning to focus on speed.",
        "R1's goal is to prove and distill reasoning (a research + transfer goal). Qwen3's goal is to ship ONE model that serves both reasoning and fast-assistant traffic (a production goal). Fusion exists because production needs both modes in one deployable unit.",
        "Qwen3 is a superior version of R1 and replaces it.",
        "Fusion is only needed because Qwen3's base is weaker than R1's."
      ],
      "answer_index": 1,
      "rationale": "The two references have different goals. R1 proves reasoning emerges (R1-Zero) and transfers cleanly (distillation) — a research + transfer thesis. Qwen3 ships a production-deployable model that does both modes — an engineering thesis. Fusion is the stage that serves the engineering goal. The strongest teams borrow from both."
    },
    {
      "id": "Q10", "bloom": "analysis", "type": "multiple_choice",
      "prompt": "Why is the 36T+ pretraining token scale load-bearing for the hybrid thesis specifically (not just for general capability)?",
      "options": [
        "Because more tokens always means a better model, regardless of architecture.",
        "Because a fused model is only a good hybrid if the non-thinking mode is a genuinely capable assistant, not a degraded reasoner. The pretraining scale is what guarantees non-thinking strength — it ensures the fast mode inherited broad capability, so fusion did not sacrifice it.",
        "Because fusion requires 36T tokens to work mechanically.",
        "Because the thinking budget only functions above 36T tokens."
      ],
      "answer_index": 1,
      "rationale": "The hybrid thesis requires BOTH modes to be strong. If the non-thinking mode is weak, the 'hybrid' is really just a reasoner with a useless fast mode. The 36T+ pretraining ensures the non-thinking mode inherits genuine broad capability from the base, so fusion adds a reasoning mode without gutting the assistant mode. This is why Qwen3 is a true hybrid, not a degraded reasoner."
    }
  ]
}