"Question"	"Answer"	"Tag"
"State the two cheaper-than-RL routes to a reasoning small model covered in FT15, and what each shares or does not share with GRPO."	"(1) CoT distillation: SFT a smaller student on chain-of-thought traces generated by a strong reasoning teacher (R1, o1, Claude/GPT with thinking). No reward model, no policy gradient, no KL anchor — pure SFT. (2) Rejection-sampling FT (RFT/RAFT): generate candidates with the current model, filter by a verifiable reward, retrain (SFT) on the winners only. Shares GRPO's rollout+reward machinery but NOT its policy-gradient training step. Both are terminal techniques, not just warmup for RL."	c3::ft15::recall
"Describe the DeepSeek-R1-Distill-Qwen-32B pipeline and its headline result."	"DeepSeek used their full R1 model (a frontier reasoner trained with full RL on verifiable rewards — the FT14 lineage) as a TEACHER. They generated reasoning traces, curated them to ~800K samples (math + code + general, verified correct), and SFT'd an off-the-shelf Qwen-32B base on those traces. NO RL on the student — just SFT. Result: R1-Distill-Qwen-32B BEATS OpenAI o1-mini on several benchmarks including AIME 2024 and MATH-500. This single fact made distillation the default, accessible recipe for small reasoning models."	c3::ft15::recall
"What is the RFT/RAFT loop, step by step?"	"(1) GENERATE: sample N candidates per problem from the current model (temperature > 0, several rollouts each). (2) FILTER by a verifiable reward: math checks? code passes tests? answer matches ground truth? Keep ONLY the positively-rewarded candidates; discard failures. (3) RETRAIN: SFT on the accepted candidates only (cross-entropy, no policy gradient). (4) REPEAT: generate with the improved model, filter, retrain. Each round trains only on the model's own successes."	c3::ft15::recall
"State the arXiv:2504.11343 finding and why it is load-bearing for this module."	"The RAFT paper found that a simple rejection-sampling fine-tuning baseline is SURPRISINGLY COMPETITIVE with GRPO and PPO on math reasoning — same base, same data, same compute. Not always equal or better, but close enough that RFT's simplicity and stability advantages make it a legitimate TERMINAL choice, not merely a warmup before 'graduating' to GRPO. This is load-bearing because it establishes RFT as a destination technique: simpler methods should not be dismissed. RFT is cheaper and more stable than full RL, and a legit standalone technique."	c3::ft15::recall
"Why is CoT distillation an instance of the FT00 thesis (steering, not teaching)?"	"The student's base already saw enormous amounts of math/code during pretraining — the reasoning capability is LATENT. Distillation does not inject new reasoning ability; it steers the student to RELIABLY EMIT a structured <think> trace and to USE the reasoning pathways the base already possesses. The teacher's trace is the steering wheel. This is why a 32B distilled model punches above its pretraining weight class: capability was latent, distillation routes probability mass through it. The corollary/limit: distillation CANNOT make the student exceed the teacher — imitation has a ceiling."	c3::ft15::analysis
"What is the structural reason RFT is more stable than GRPO?"	"The training step. GRPO's policy gradient is advantage × log-prob summed over rollouts — it learns from BOTH successes and failures, and its failure modes (rollout collapse, KL drift, flat-reward noise) are all POLICY-GRADIENT pathologies that can silently ruin the model while the reward looks fine. RFT's training step is SFT — cross-entropy on accepted candidates only. SFT does not diverge like policy gradients. RFT's worst cases (small accepted set, wrong reward) are DIAGNOSABLE data-quality problems you can inspect by reading what you accepted. RFT trades expressivity (ignores failures) for stability + diagnosability."	c3::ft15::analysis
"State the three-way decision (distillation vs RFT vs GRPO) and the three axes that determine it."	"Axis 1 — Strong teacher available? Yes → distillation on the table (cheapest, consider first). No → must generate own signal (RFT or GRPO). Axis 2 — Reward verifiable (math, code, logic w/ checks)? Yes → RFT & GRPO available. No → neither applies (FT13 preference territory). Axis 3 — Compute/stability: constrained budget / must-not-diverge → RFT; large budget / willing to tune KL / push frontier → GRPO. In one line: distillation = cheapest (have teacher, don't need to exceed it); RFT = middle (own data, reward filter, no RL instability); GRPO = most powerful/expensive (push beyond both)."	c3::ft15::application
"Why does distillation have a ceiling at the teacher's quality, and what does that imply for technique choice?"	"Distillation is imitation: the student learns to reproduce the teacher's <think> traces. The student's reasoning ceiling is the teacher's ceiling minus whatever the smaller model forgets. You cannot distill reasoning the teacher does not have. Implication: if your goal is to push BEYOND what the teacher can produce, distillation is insufficient — you need RL on the student's OWN rollouts (RFT or GRPO), where the model can discover solutions the teacher never generated. Distillation is right when you want a cheap, small, reliable reasoner and do not need to exceed the teacher."	c3::ft15::analysis
"What is a hybrid model (thinking toggle), and what problem does it solve?"	"A single model that can run in THINKING mode (emit <think> trace, then answer — high accuracy, high latency) or NON-THINKING mode (answer directly — low latency), toggled per request via a control token or system instruction. Qwen3 is the canonical example. It solves the LATENCY TAX: a reasoning model that ALWAYS emits a long think block is slow on trivial queries. The toggle is NOT a separate model — same weights, steered by a flag. One deployed model serves both the cheap-fast and expensive-accurate paths."	c3::ft15::recall
"What is a thinking budget, and how does it differ from a binary thinking toggle?"	"A thinking budget is a mechanism for ADAPTIVE COMPUTE: rather than a binary on/off toggle, the model is given a token budget for thinking — think for up to N tokens, then answer. This lets the model spend more compute on hard problems (large N) and less on easy ones (small N), allocating inference cost to where it buys accuracy. It makes the thinking/non-thinking choice continuous rather than binary. Implication for FT: a hybrid student trains on a MIX of thinking and non-thinking traces to learn both modes and the ability to choose."	c3::ft15::recall
"Name the three anti-patterns FT15 warns against and the fix for each."	"(1) Distilling from a weak teacher — garbage in, garbage out; the ceiling is the teacher's. Fix: use a genuinely strong teacher on the target domain (R1 worked because R1 is frontier). (2) RFT without reward verification — a noisy reward means you train on incorrectly-accepted candidates. Fix: use a VERIFIABLE reward (math checks, strict tests); preference-style rewards are a different regime (FT13). (3) Ignoring RFT's stability advantage over GRPO — reaching for GRPO by default. Fix: try RFT first when the reward is verifiable; escalate to GRPO only when RFT plateaus with budget remaining."	c3::ft15::application
"Why is RFT described as 'same data-generation machinery as GRPO, radically different training step'?"	"Both use rollouts from the current model and a verifiable reward to score them — that is the shared data-generation machinery. The difference is the TRAINING STEP. GRPO uses the rollouts to compute a policy gradient (advantage × log-prob) — it learns from successes AND failures, pushing up good and down bad. RFT DISCARDS the failures and supervises (SFT) on the successes only. Same rollouts and reward; GRPO keeps the gradient signal from all samples, RFT throws away the failures and trains only on winners. This is why RFT has SFT stability while GRPO has policy-gradient instability."	c3::ft15::analysis
"What does it mean that distillation and RFT are 'terminal techniques, not just warmup steps'?"	"A terminal technique is one you can ship from — a destination, not merely a stage on the way to something else. R1-Distill-Qwen-32B is a SHIPPING model trained with distillation only (no RL on the student). The arXiv:2504.11343 result shows RFT is competitive with GRPO on math reasoning, so RFT too can be the final technique. The anti-pattern is treating distillation/RFT as merely pre-RL initialization — you then spend RL compute for what may be marginal gains over a strong distilled/RFT'd baseline. Distill (or RFT), evaluate against your target, escalate to GRPO ONLY if it falls short."	c3::ft15::analysis
"What is the defensible stack that combines distillation, RFT, and GRPO, and why is each stage optional?"	"Distill first (cheap, gets you most of the way using a strong teacher's traces) → RFT to self-improve on areas the teacher was weak (generate own candidates, filter by reward, retrain on winners) → GRPO ONLY if RFT plateaus and budget remains (push the frontier with full RL). Each stage is optional: many production reasoning models STOP at distillation (R1-Distill-Qwen-32B is shipped, not escalated). Escalation is justified only by measured shortfall against the target, not by default. This avoids paying full RL cost for gains cheaper techniques already capture."	c3::ft15::application
"A team wants a small math-reasoning model. They have a frontier API with thinking and a constrained budget. Which technique, and why?"	"DISTILLATION. They have a strong teacher (the frontier API with thinking) and a constrained budget. Distillation is the cheapest path that works at scale: generate <think> traces from the API on math problems, curate (verify answers), SFT the small student. No RL, no policy-gradient instability, SFT compute profile. The ceiling is the teacher's — but for a constrained-budget small model, matching a frontier teacher is already a win (cf. R1-Distill-Qwen-32B beating o1-mini). RFT/GRPO would be escalation if the distilled model falls short AND budget later allows."	c3::ft15::application
"A team wants to push a reasoning model PAST what any available teacher can do, on math. They have a verifiable reward and a large budget. Which technique, and why?"	"GRPO (FT14). The goal is to EXCEED the teacher — distillation is capped at the teacher's ceiling. The reward is verifiable (math) and the budget is large, so the conditions for full RL are met. RFT could be tried first (cheaper, stable, competitive per 2504.11343) as a stepping stone, but the terminal goal of pushing past the teacher requires learning from the model's OWN rollouts via policy gradient — which is GRPO's domain. RFT's limit is that it ignores failures; GRPO learns from them, which is the edge on the hardest exploration problems."	c3::ft15::application
"A team runs RFT with a fuzzy LLM-judge reward on open-ended writing. Diagnose the problem."	"The reward is NOT verifiable. RFT's entire signal is the FILTER — if the reward is a noisy LLM judge with false positives, the team trains on INCORRECTLY-ACCEPTED candidates, producing a model that emits plausible-looking wrong/unrewarding outputs. RFT demands a verifiable reward (math checks, code passes strict tests, ground-truth match). An open-ended writing preference is a PREFERENCE problem — that is FT13 (DPO family) territory, not reasoning FT. The fix is either to switch to a verifiable reward (if one exists for the task) or move to preference optimization (DPO/KTO), not RFT with a fuzzy reward."	c3::ft15::analysis
"Why did R1-Distill-Qwen-32B beat o1-mini despite being trained with SFT only (no student-side RL)?"	"Because (1) the teacher (R1) is a frontier reasoner whose traces encode high-quality reasoning — the student imitates genuinely good chains, not mediocre ones; (2) the curation was large and high-quality (~800K verified-correct samples across math/code/general); (3) the Qwen-32B base already had latent math/code capability from pretraining — distillation ROUTED probability mass through those pathways reliably and in a structured trace (FT00 steering). The base was capable; the traces made it reliable. This shows distillation's ceiling-is-teacher property is not a weakness when the teacher is frontier: matching a frontier teacher at 32B already beats a compressed-reasoning product."	c3::ft15::analysis
"How does the training compute profile of CoT distillation compare to GRPO, and why?"	"Distillation's compute profile is that of ORDINARY SFT — a single forward+backward pass over the (problem, trace, answer) tuples, no rollouts at train time, no reward model, no KL anchor, no reference model. GRPO's profile is a full RL loop: generate K rollouts per problem per step (expensive inference), score with the reward, compute advantages, policy-gradient update with KL to a reference. The teacher in distillation did the expensive search ONCE at data-gen time; the student just imitates. So distillation is dramatically cheaper per training step and far simpler to run. This is why it is the accessible default and GRPO is the escalation."	c3::ft15::analysis
"What is the difference between a teacher used for distillation and a reward used for RFT/GRPO?"	"A TEACHER provides TARGET DATA — concrete <think> traces the student imitates via SFT. It is a generative source of correct reasoning. A REWARD provides a SCALAR SIGNAL — a score (pass/fail, correct/incorrect) that filters or weights candidates. It does not provide the trace; it only judges. Distillation needs a teacher (you copy its traces). RFT/GRPO need a reward (you generate your own candidates and judge them). You can have a reward without a teacher (verifiable math — no frontier model needed). You can have a teacher without a reward (distillation does not score the student's outputs during training). RFT/GRPO cannot use a teacher as their primary signal — they need a verifiable reward to filter/score."	c3::ft15::analysis
"Why might distilling traces that are 'too long, too idiosyncratic, or poorly tokenizer-fit' produce a degraded student?"	"Because distillation is IMITATION, and imitation requires the target to be imitable. If the teacher's traces are excessively long, the small student cannot reliably reproduce them (capacity limit) and truncates or drifts. If they are stylistically idiosyncratic (unusual tokens, erratic formatting), the student spends capacity mimicking surface style rather than reasoning structure. If the tokenizer handles the trace's tokens poorly (inefficient tokenization, unknown tokens), the student's representation of the trace is degraded. Fix: curate the traces — trim to reasonable length, normalize formatting, ensure the format is something the student can reproduce. The target text must be learnable by the student, not just correct."	c3::ft15::application
"A colleague says: 'Distillation is just a warmup for RL — real reasoning models need GRPO.' Evaluate against FT15."	"Wrong, per FT15. R1-Distill-Qwen-32B is a SHIPPING model trained with distillation only (no student-side RL) that beats o1-mini. Distillation is a TERMINAL technique, not merely an init step. The arXiv:2504.11343 result shows the same for RFT — competitive with GRPO. Treating distillation as only a warmup leads to spending RL compute for what may be marginal gains over a strong distilled baseline. The correct framing: distill, EVALUATE the distilled model against your target, and escalate to RL ONLY if it falls short. Many production reasoning models stop at distillation. GRPO is for pushing past the teacher/plateau, not a default requirement."	c3::ft15::analysis
"What does it mean that RFT 'trades expressivity for diagnosability' compared to GRPO?"	"EXPRESSIVITY: RFT trains only on ACCEPTED (positively-rewarded) candidates — it ignores failures, so it cannot learn the 'push down the bad' signal that GRPO's policy gradient encodes. On the hardest exploration problems (where learning from failures matters), GRPO has an edge. DIAGNOSABILITY: in RFT, what you train on is EXACTLY what you accepted — you can inspect the accepted set and see if the reward is wrong or the set is too small. Failures are data-quality problems you can read. In GRPO, a silent mode collapse can ruin the model while the reward looks fine — the failure is in the gradient dynamics, invisible without training-metric forensics. RFT gives up some learning signal for transparency."	c3::ft15::analysis