"Question"	"Answer"	"Tag"
"What is the course stance on studying uncensored lineages like Dolphin and Hermes?"	"Studied as ENGINEERING CASE STUDIES, not advocacy. The stance (FT00) is unchanged and absolute: the model steers, the harness bounds. An uncensored model is only responsible inside an eval'd harness. Uncensoring RAISES the harness requirement — it does not lower it. Both extremes (celebrate / condemn) miss the engineering."	c3::ftdd06::recall
"What is Hermes 3 (Nous Research), and what is its recipe per arXiv:2408.11857?"	"Hermes 3 is a full-parameter SFT+DPO steer of Llama 3.1 at 8B/70B/405B, 'neutrally-aligned, unlocked, uncensored, highly steerable,' trained on primarily synthetic responses. Recipe: (1) full-param SFT (format/instructions/character), then (2) DPO (sharpen toward compliance + steerability). It is the canonical large-scale alignment-control recipe in open weights."	c3::ftdd06::recall
"What is OpenHermes 2.5, and why is it significant to the course thesis?"	"Teknium's instruction dataset of ~1,001,551 examples (~1M), open-source + custom synthetic. It is the steering wheel behind the OpenHermes 2.5, Nous Hermes 2, and Hermes 3 model families. It makes the FT00 thesis concrete: 'data matters more than algorithm' — one high-quality dataset steered a whole family. (Flaws propagate too; audit data like code.)"	c3::ftdd06::recall
"Who created the Dolphin series, and what is its design philosophy?"	"Eric Hartford (Cognitive Computations). Philosophy: COMPLIANCE OVER JUDGMENT — the model complies with user instructions and deploying-system policy rather than imposing judgmental refusals. Executed primarily via DATASET CURATION (train toward compliance), not primarily abliteration. The Dolphin lineage spans Llama and Mistral families."	c3::ftdd06::recall
"What makes Dolphin3.0-R1-Mistral-24B technically distinctive?"	"It is the ONLY uncensored model trained on DeepSeek-R1 reasoning traces. Base: Mistral Small 24B (Instruct-2501). Trained over 3 rounds on ~800K reasoning traces from the Dolphin-R1 dataset (distilled from R1). It pairs Hartford's compliance-over-judgment philosophy with R1-grade reasoning. Sits at the intersection of FT14 (reasoning), FT15 (CoT distillation), FT16-18 (alignment control)."	c3::ftdd06::recall
"Why does Hermes 3 use full-parameter fine-tuning instead of LoRA, and what is the cost/benefit?"	"Full-param finds a HIGHER-RANK solution (per FT00 evidence) that can make larger behavioral adjustments — justified because Hermes 3 substantially shifts the model's character (toward neutrality/steerability). Cost: full FT of 405B is a serious cluster job. Benefit: the higher-rank path moves behavior further than a light LoRA steer would. For a light format steer, LoRA would suffice and be far cheaper."	c3::ftdd06::application
"Map the two Hermes 3 stages onto the Steering Stack and FT modules."	"Stage 1 (full-param SFT) = Layer 3 (the Steer) -> FT12; steers format/instruction-following/character. Stage 2 (DPO) = Layer 3 -> FT13; sharpens preference toward compliance + steerability ('unlocked, highly steerable'). The base is Layer 1 (Llama 3.1, open-weights-only per FT02). Layer 5 (the harness) is the DEPLOYER's responsibility, not Hermes 3's."	c3::ftdd06::application
"What is the measurable capability cost of uncensoring (FT17), and what causes it?"	"GSM8K math scores drop from +1.5pp (rare gain) to -18.8pp depending on tool/model (Dec 2025 study). Cause: the refusal direction in the residual stream is ENTANGLED with other capabilities. Whether removed by abliteration (weight edit) or data-driven compliance steering (training), you nudge the entangled capabilities. An uncensored model is a different point in capability-compliance space, not 'same model minus refusals.'"	c3::ftdd06::application
"Distinguish Hartford's Dolphin approach from abliteration (FT17)."	"Dolphin = DATA-DRIVEN COMPLIANCE STEERING: train toward compliance from the dataset up (filter refusal-inducing patterns, curate compliance-oriented examples). Abliteration = POST-HOC WEIGHT EDIT: delete a refusal direction in the residual stream. Different Layer 3 operations with different cost/reproducibility profiles. Dolphin lineages sometimes combine both, but Hartford's approach is primarily data-driven."	c3::ftdd06::application
"Restate the FT00/FT23 synthesis using Dolphin and Hermes 3 as concrete examples."	"'Uncensor the model so it executes; harness the model so it executes only what it should.' Dolphin3.0-R1 is steered to execute (compliance + R1 reasoning) — only responsible inside a harness that bounds what it MAY do. Hermes 3 is steered to be highly steerable — its steerability means a weak harness does damage. The harness requirement is RAISED by uncensoring, not lowered."	c3::ftdd06::analysis
"Why is deploying an uncensored model without an eval'd harness the cardinal deployment error for this module?"	"A compliance-oriented model will not self-refuse — it executes. Without a harness providing policy gates, audit logs, and a threat model, it executes the dangerous things too. An uncensored model in a WEAK harness is strictly MORE dangerous than a refusal-trained model in a weak harness, because the refusal-trained model at least self-gates some harmful requests. The harness is what makes the uncensored model responsible."	c3::ftdd06::analysis
"How does OpenHermes 2.5 being the shared dataset backbone create both a strength and a liability?"	"Strength: one high-quality ~1M-example dataset steered a whole family (OpenHermes 2.5 -> Nous Hermes 2 -> Hermes 3) — data matters more than algorithm (FT00). Liability: any flaw (bias, quality regression) in the synthetic-data pipeline propagates into EVERY downstream model. The steering wheel's flaws steer you into the wall, precisely and at scale. Audit your data like code."	c3::ftdd06::analysis
"Why is Hermes 3's 'high steerability' a double-edged property?"	"Steerability means the deploying system (prompt + harness) has FINE control — a feature for legitimate use. But the same steerability means a WEAK harness produces a model that does whatever it is prompted to do, including the dangerous things. The model gives you control; it does not give you safety bounds. Safety bounds are the harness's job (Layer 5). High steerability RAISES the harness requirement."	c3::ftdd06::analysis
"A team picks Dolphin3.0-R1 for a math-heavy agentic task and is surprised math quality dropped. What did they likely miss, and what's the fix?"	"They read the strengths in the model card and ignored the capability-cost trade-off (FT17: up to -18.8pp GSM8K). An uncensored model is a DIFFERENT point in capability-compliance space, not 'same model + compliance.' Fix: read the FULL eval table including regressions; if math is load-bearing, a refusal-trained model behind a good harness policy may serve better. Decide on numbers, not ideology."	c3::ftdd06::analysis
"Why are Dolphin/Hermes described as the 'best-documented large-scale examples of the FT16-FT18 techniques'?"	"They are full-parameter alignment-control steers of major bases (Llama 3.1, Mistral) at scale (up to 405B), with published recipes (Hermes 3 has arXiv:2408.11857; Dolphin ships model cards). They demonstrate SFT (FT12) + DPO (FT13) + compliance/uncensoring (FT16-18) + reasoning distillation (FT14-15) as PRODUCTION runs, not toy examples. Few open-weights lineages are documented this thoroughly."	c3::ftdd06::analysis
"Explain why 'compliance over judgment' makes a model MORE harness-dependent, not less."	"A judgment-oriented model self-refuses some harmful requests (a partial internal gate). A compliance-oriented model defers to the user + harness and will NOT self-refuse — so every gate the judgment model provided internally must now be provided EXTERNALLY by the harness. Removing the model's internal judgment shifts the entire safety burden onto the harness. Less internal gating = more external gating required."	c3::ftdd06::analysis
"Connect Dolphin3.0-R1-Mistral-24B's recipe to three FT modules and explain the intersection."	"(1) FT14/FT15: it distills ~800K DeepSeek-R1 reasoning traces over 3 rounds (CoT distillation from a stronger reasoner). (2) FT12/FT13: SFT/DPO for format and character. (3) FT16-18: compliance-over-judgment uncensoring via dataset curation. The intersection: it is a single production model demonstrating reasoning (Pillar 4) + alignment control (Pillar 5) simultaneously — the synthesis those pillars aim for."	c3::ftdd06::analysis