"What is DeepSeek-R1's role in this course?"	"The canonical reasoning distillation reference. It is the strongest published evidence that reasoning is a steering problem, not a knowledge problem. R1-Zero proves emergence from RL; R1 is the polished pipeline; the distillation (SFT-only) made CoT distillation standard."	c3::ftdd07::recall
"What base model did R1-Zero use, and what was its key property for the experiment?"	"DeepSeek-V3-Base — a 671B Mixture-of-Experts model (37B active per token). The key property: NO instruction tuning, NO supervised reasoning data. A standard pretrained base. This is what makes the emergence result meaningful."	c3::ftdd07::recall
"What reward signal did R1-Zero use, and what is significant about this choice?"	"Verifiable rewards ONLY: math correctness, code execution pass/fail, LeetCode verdicts. No reward model, no judge, no human preferences. The significance: if you can verify, you don't approximate with a learned judge. This became a field-wide lesson."	c3::ftdd07::recall
"What emerged from R1-Zero's RL training, and why is it significant?"	"Long, structured chain-of-thought emerged unprompted: self-correction ('wait, let me reconsider'), re-derivation of formulas, answer substitution checks, the 'aha moment.' Significant because nobody put reasoning demonstrations in the data — the reward made deliberation pay off and deliberation grew."	c3::ftdd07::recall
"Why is R1-Zero the strongest evidence for the course thesis (reasoning is steering)?"	"If reasoning were a knowledge problem, GRPO on a base with zero reasoning traces could not produce it — you'd need demonstrations or continued pretraining. It IS steering: the base already saw billions of reasoning tokens in pretraining. RL redirects probability mass so reasoning pathways become the chosen behavior. That is what happened."	c3::ftdd07::analysis
"Why was R1-Zero 'a proof, not a product'?"	"It was messy: mixed languages mid-chain (Chinese/English tangled), run-on chains that were correct but ugly, sometimes failed to converge to a clean final answer despite correct intermediate work. Benchmark-competitive but not shippable. R1 was built to make the emergence usable."	c3::ftdd07::recall
"Name the four stages of the R1 pipeline in order."	"(1) Cold-start SFT — anchors format with thousands of curated long-CoT examples. (2) Reasoning RL (GRPO) — verifiable rewards, scales capability. (3) Rejection-sampling SFT — generate many candidates, filter correct + readable, ~800K clean traces. (4) Final RL — whole distribution, aligns the reasoner into a product."	c3::ftdd07::recall
"What does Stage 1 (cold-start SFT) fix, and why is it necessary?"	"It fixes R1-Zero's messiness (unreadable chains, language mixing). It anchors tone, a single language, a clean <think> structure, and a properly formatted final answer. It does NOT teach reasoning — the base has that. It steers the model into the output shape RL will amplify. Small in data volume, large in consequence."	c3::ftdd07::application
"What is rejection sampling (Stage 3) and why is it the load-bearing stage for distillation?"	"Generate many candidate chains per problem with the Stage-2 RL'd model, filter for correct final answers, then filter for quality (no language mixing, readable). This produces a large, clean SFT dataset that reliably reproduces the RL'd model's BEST behavior. This curated trace set (~800K) is the raw material for the distillation into the six dense students."	c3::ftdd07::application
"What distinguishes Stage 2 RL from Stage 4 RL in the R1 pipeline?"	"Stage 2 optimizes ONLY reasoning tasks (verifiable rewards: math, code, logic). Stage 4 runs a SECOND RL pass across the FULL distribution — reasoning AND general assistant tasks — with both verifiable reasoning rewards and preference/policy signals. Stage 2 scales capability; Stage 4 turns the strong reasoner into a shippable product."	c3::ftdd07::analysis
"How many dense models were distilled from R1, and what was the training method?"	"SIX dense models: Qwen-1.5B, Qwen-7B, Qwen-14B, Qwen-32B, Llama-8B, Llama-70B. Training method was SFT-ONLY on ~800K curated reasoning traces (plus non-reasoning SFT data). No GRPO, no reward model, no RL loop on the students. Just supervised fine-tuning on the teacher's best traces."	c3::ftdd07::recall
"What result made CoT distillation standard, and why?"	"R1-Distill-Qwen-32B beat OpenAI o1-mini on AIME 2024 (competition math). A 32B dense model trained by SFT only outperformed a frontier proprietary reasoning model. This showed you no longer needed the RL supercomputer — just a strong teacher's traces and a standard SFT run. After this, CoT distillation became the default way to build a reasoning model."	c3::ftdd07::analysis
"Why did RL on top of the distilled students give only marginal gains?"	"The teacher's curated traces were already a near-optimal steering signal — they showed the student the high-reward trajectories directly. RL on top could only refine what the traces had installed, and the marginal refinement was small relative to compute cost. Takeaway: distill first; RL the student only with a verifiable-reward domain the traces under-cover."	c3::ftdd07::analysis
"How does the distillation result support the steering thesis from the other direction?"	"If reasoning were a capability the small models LACKED (a knowledge gap), distilling traces could not transfer it. The fact that SFT on traces DID transfer reasoning is evidence the small bases already had latent capability. The traces STEERED them to express it. The student inherits the BEHAVIOR, not new ABILITY."	c3::ftdd07::analysis
"What are the three things the field took from R1?"	"(1) CoT distillation is the practical path to a reasoning model (before R1: assume you need full RL; after: teacher traces + SFT). (2) Verifiable rewards beat reward models for reasoning RL (if you can verify, don't approximate with a judge). (3) RL-on-base proves the capability is latent (R1-Zero is the reference everyone cites for 'reasoning is steering')."	c3::ftdd07::recall
"How do R1 and Qwen3 differ as reasoning references?"	"R1 = the DISTILLATION reference: teacher pipeline + SFT-only transfer. Use when building a reasoner by distilling from a teacher. Qwen3 = the HYBRID reference: one model fusing thinking + non-thinking modes with a thinking-budget mechanism. Use when shipping a production model that adapts compute to the question. They prove different things; know both."	c3::ftdd07::analysis
"For a team with a strong reasoning teacher and limited compute, what is the correct first step to build a reasoning student, per R1's evidence?"	"SFT-only distillation from the teacher's curated traces. R1's own ablation shows SFT-only captures most of the gain and RL on the student is marginal. Reach for GRPO only if there is a verifiable-reward domain the traces under-cover and the compute to spend. Distill first."	c3::ftdd07::application
"Where is R1 published and where are the weights?"	"Published as arXiv:2501.12948 (January 2025), then in Nature (s41586-025-09422-z) — the peer-reviewed account. Weights open under the DeepSeek License at github.com/deepseek-ai/deepseek-r1, including the distilled students. The Nature publication gave the field a citable, refereed source for the emergence result."	c3::ftdd07::recall
"Using verifiable rewards (execution checks) instead of a reward model for RL — what is the principle, and when does it apply?"	"Principle: if you can verify the answer by checking correctness (math, code execution, LeetCode), do not approximate the verification with a learned judge. R1's reasoning stages used only verifiable rewards. It applies when the task has a ground-truth check; for subjective tasks (writing quality) you still need a preference signal."	c3::ftdd07::application
"Why is distilling from a weak teacher a fundamental error, regardless of distillation method?"	"The teacher's quality is the CEILING. A weak or misaligned teacher produces traces that steer the student toward mediocre reasoning. The student learns to imitate the teacher's behavior, including its flaws. Garbage traces in, mediocre reasoner out. No amount of distillation sophistication recovers from a teacher that cannot itself produce high-reward trajectories."	c3::ftdd07::analysis