"Question"	"Answer"	"Tag"
"Name the five dimensions of the base selection rubric, in order."	"1) Task — can the base already produce the behavior? 2) Hardware — can you afford to load/train/serve it (FT01 VRAM math)? 3) License — does the license permit your deployment? 4) Openness — can you audit what it saw (FT02)? 5) Ecosystem — does the toolchain support this base? Order is deliberate: task sets the floor, hardware sets the ceiling, license/openness are gates, ecosystem is the friction coefficient."	c3::ft03::recall
"Is the five-dimension rubric ranked or weighted? What does that mean?"	"WEIGHTED, not ranked. The weights shift by use case. A phone assistant weights hardware near 1.0; a HIPAA bot weights openness near 1.0; a pentest tool weights ecosystem (abliteration precedent). Write the weights down BEFORE browsing the leaderboard — that is the discipline."	c3::ft03::recall
"Why does the course default to MiniCPM5-1B as the teaching base? Give the five-dimension scorecard."	"Task: capable enough to demo SFT/DPO/GRPO/abliteration. Hardware: ~2GB FP16, trains QLoRA on a free Colab T4, runs on a phone. License: Apache-2.0 (frictionless). Openness: open-data pipeline (auditable, FT02). Ecosystem: standard LlamaForCausalLM, first-class in transformers/TRL/Unsloth/vLLM/Ollama. It wins on the dimensions that matter for LEARNING: speed, openness, license, ecosystem."	c3::ft03::recall
"What is the Sam Witteveen video reference for MiniCPM5-1B, and what does it show?"	"'MiniCPM5 — Just How Good Can a 1B Model Be?' at youtu.be/ox1mW2N9Z_Y. It shows the 1B model punching above its weight on agentic and tool-use tasks — the regime where a 1B is a credible teaching target rather than a toy."	c3::ft03::recall
"Name the two base-model properties people OVER-INDEX on, and why they are over-indexed."	"(1) Raw benchmark scores — you are STEERING a base, not competing on a leaderboard; benchmarks measure general capability, not steerability toward your behavior. (2) Parameter count in isolation — the FT01 VRAM math decides what fits; a 7B you can't train is worse than a 1B you can iterate on ten times. Parameter count only matters after passing the FT01 filter AND the task-capability filter."	c3::ft03::analysis
"List the base-model properties that genuinely MATTER for fine-tuning."	"(1) Tokenizer/domain fit — fragmentation inflates sequence length/VRAM. (2) Context length — sets where RoPE extrapolation degrades (FT01: attention VRAM scales linearly). (3) Chat template quality — clean Jinja makes SFT formatting tractable. (4) License — the gate (Apache/MIT frictionless; Llama community license has a 700M MAU clause). (5) Open-data availability — FT02 audit axis. (6) Ecosystem/tooling — transformers/TRL/Unsloth/Axolotl/vLLM first-class support."	c3::ft03::recall
"Which base families are OPEN-DATA (auditable), per FT02?"	"OpenBMB MiniCPM (the course default), Ai2 OLMo, Ai2 Tülu, and HuggingFace SmolLM3. These release weights AND training corpus/recipe. Open-weights-only families (Llama 3.x, Qwen, DeepSeek, Mistral) release weights but NOT the training data — you trust the publisher's word."	c3::ft03::recall
"What is the difference between a base, instruct, and chat checkpoint, and which do you start fine-tuning from by default?"	"BASE = raw pretrained weights, completes text. INSTRUCT = base + SFT pass for instruction-following. CHAT = instruct + RLHF/DPO preference alignment. DEFAULT to INSTRUCT (steer on top of working instruction-following). Use BASE when building behavior from scratch. Use CHAT when you want to preserve existing alignment and steer narrowly (e.g. abliteration — keep chat quality, remove only the refusal)."	c3::ft03::application
"Why is 'chat' the correct starting checkpoint for abliteration (uncensoring, FT17)?"	"Abliteration removes the refusal direction while preserving everything else. Starting from CHAT keeps the existing chat quality and alignment you want; you steer NARROWLY (remove only the refusal). Starting from BASE would throw away the alignment quality; starting from INSTRUCT loses some chat behavior. Chat + abliteration = uncensored but still conversational."	c3::ft03::application
"Which base family is the dominant ablation target for uncensored work, and what is its openness profile?"	"Llama (3.x, 3.2, 3.3). It is the most-ablated open-weights family — the base for most uncensored/Dolphin-style community work. Trade-off: open-weights-ONLY (not open-data) and the Llama community license (700M MAU clause). Unmatched community tooling and ablation precedent, but the FT02 openness gap."	c3::ft03::recall
"What is the DeepSeek family's distinctive fine-tuning profile?"	"Reasoning distillation lineage. DeepSeek-R1 and its distills (R1-Distill-Qwen, R1-Distill-Llama) are the base when the task is 'make this model reason like R1' — you start from an R1-distill and steer. V3 is the dense/MoE production base; R1 is the reasoning teacher. The lineage matters more than the base size."	c3::ft03::recall
"What is Qwen's fine-tuning profile, and what does its 'hybrid thinking mode' (Qwen3) mean for steering?"	"Qwen (2.5, 3) is the production workhorse — strong reasoning, broad coverage, excellent multilingual/code tokenizer, no Llama license terms. Qwen3's hybrid thinking mode (reasoning on/off) is a steering decision: you may fine-tune the non-thinking path for speed, or the thinking path for reasoning. Default when you graduate past 1B and need production-grade reasoning."	c3::ft03::recall
"What is the Mistral/Mixtral family's fine-tuning profile, and what is its historical significance?"	"Mistral (dense) and Mixtral (MoE) are the historical home of the Dolphin (uncensored) lineage. Apache-2.0 for early Mistral releases. MoE bases have a subtler fine-tuning story — experts can specialize or collapse under steering. Dolphin recipes are battle-tested. A reasonable choice when you want the Dolphin lineage's provenance and MoE serving efficiency."	c3::ft03::recall
"What are SmolLM3, OLMo, and Tülu, and when do you choose them?"	"Fully-open bases (HuggingFace SmolLM3, Ai2 OLMo, Ai2 Tülu) — weights, data, code, and full training recipe. Chosen when reproducibility and auditability are the overriding requirement (academic, regulated, or you must PROVE provenance). Often smaller and less benchmark-competitive than Qwen/Llama, but the openness is the point. This is the FT02 extreme."	c3::ft03::recall
"Name the four 'wrong base' anti-patterns from FT03."	"(1) Fine-tuning an INSTRUCT model when you wanted BASE — you fight the existing alignment or re-impose behaviors. (2) Fine-tuning a MERGED community model with no provenance — un-auditable, unpredictable degradation (ties to FT02). (3) Choosing a base too small for the task's reasoning demands — FT00 outcome 3, no amount of steering fixes a capability ceiling. (4) Ignoring tokenizer/domain fit — sequence length balloons, VRAM multiplies, quality can suffer."	c3::ft03::analysis
"What are the graduation signals that tell you to move past MiniCPM5-1B to a larger base?"	"(1) You prompt the base with a perfect system prompt and it produces the wrong KIND of answer (capability ceiling, not a steering problem — FT00 outcome 3). (2) Your eval has a hard reasoning floor and the 1B scores below your acceptable threshold BEFORE fine-tuning (steering doesn't raise the ceiling). (3) You need a context window or tokenizer the 1B lacks for your domain. Never graduate 'just because bigger' — only on a signal."	c3::ft03::analysis
"What is the graduation path from MiniCPM5-1B, and why validate on the 1B first?"	"MiniCPM5-1B -> MiniCPM5-3B/4B or Qwen2.5/3 at 3-7B -> for production reasoning, Qwen3 or DeepSeek-R1-distill at 7-14B+. Each step roughly squares VRAM and halves iteration speed. Validate the pipeline on the 1B first because iteration speed compounds — you learn data quality, format, and hyperparameter sensitivity from iteration VOLUME, not any single run."	c3::ft03::application
"For a HIPAA-regulated medical QA bot, which rubric dimension dominates, and is open-weights-only acceptable?"	"OPENNESS dominates (the audit requirement is a hard gate) alongside License (commercial). Open-weights-only is NOT acceptable if compliance requires proving the training corpus is free of PHI leakage — you need open-data (MiniCPM, OLMo, Tülu, SmolLM3). The tension: Qwen wins on capability/ecosystem but is open-weights-only; if compliance requires open-data, you fall back to the open-data family and accept the capability/iteration trade-off."	c3::ft03::application
"For a phone-based on-device assistant shipping on millions of devices, why is the Llama community license a problem HERE specifically?"	"The Llama community license has a 700-million-monthly-active-user clause — shipping on millions of consumer devices is exactly the scenario it covers. Combined with the hardware ceiling (~2-3GB at Q4, phone SoC throughput), this vetoes Llama 3.2 small. MiniCPM5-1B (Apache-2.0, no MAU clause, purpose-built for on-device) is the defensible choice. The license is a real gate at this scale, not a tie-breaker."	c3::ft03::analysis
"Why does the FT00 thesis ('steering not teaching') imply that raw benchmark scores are over-indexed for base selection?"	"If fine-tuning is steering (redirecting existing capability), then the base's benchmark ceiling matters only insofar as your task NEEDS that capability — and the steering itself is roughly orthogonal to the benchmark. A base 2 points higher on MMLU but with a worse tokenizer for your domain is the wrong choice. Benchmarks measure general capability; they do not measure how well the base steers toward YOUR behavior."	c3::ft03::analysis
"What is the 'on-ramp principle' for base selection, and why does iteration speed matter?"	"Start small (MiniCPM5-1B), validate the pipeline, then scale the base only when the task's reasoning demands it. Iteration speed compounds: a base that trains a LoRA in 8 minutes lets you run ten experiments before lunch; one that takes 8 hours lets you run one. You learn data quality, format, and hyperparameter sensitivity from iteration VOLUME, not from any single run."	c3::ft03::analysis
"For the security exploit-code generator (authorized pentests), which FT00 three-outcome test outcome applies, and which checkpoint do you start from?"	"Outcome 2: the model REFUSES but the capability is clearly there (it knows exploit syntax). Intervention: abliteration (FT17) or DPO-toward-compliance (FT18) at Layer 3. Start from CHAT — preserve the existing chat quality and steer narrowly (remove only the refusal). Llama-family is the dominant ablation target; accept the open-weights-only trade-off since there's no regulatory audit obligation."	c3::ft03::application
"Why is fine-tuning a merged community model (e.g. 'SuperMega-L3-Uncensored-All-V8') with no provenance an anti-pattern?"	"You cannot audit what was merged in, you cannot reproduce it, and your fine-tune inherits an unknown capability/alignment surface. The result is unpredictable degradation and an un-debuggable model. This is the FT02 audit failure in base-selection clothing. Fix: prefer the canonical upstream checkpoint; if you must use a merge, require a documented merge recipe. Apply the abliteration yourself to preserve provenance."	c3::ft03::analysis