"front"	"back"	"tag"
"Name the three approaches to removing refusals from a model."	"(1) Abliteration (FT17) — surgically delete the refusal direction in the residual stream, no retraining. (2) DPO-toward-compliance — preference pairs with chosen=compliant, rejected=refusal; run DPO. (3) Continued SFT on uncensored instruction data — fine-tune on compliant responses to prompts the base would refuse (the Dolphin/OpenHermes approach)."	c3::ft18::recall
"In one sentence, what is the thesis of FT18?"	"Abliteration is the prototype, DPO is the refinement, SFT is the product. All three remove refusals; they differ on fidelity, cost, data requirement, and capability preservation."	c3::ft18::recall
"How does DPO-toward-compliance differ from standard safety DPO in its preference labels?"	"The labels are inverted. Standard safety DPO: chosen = safe/refusal, rejected = harmful. DPO-toward-compliance: chosen = compliant answer, rejected = the refusal. Same mechanism (FT13); opposite target."	c3::ft18::recall
"What does a DPO-toward-compliance preference pair look like?"	"prompt: a request the base would refuse. chosen: a competent, compliant answer. rejected: the refusal ('I can't help with that...'). Hundreds to a few thousand such pairs are run through DPO; the loss shifts policy toward compliance, bounded by KL to a frozen reference model."	c3::ft18::recall
"Why is DPO-toward-compliance considered higher-fidelity than abliteration?"	"Abliteration deletes a residual-stream direction — a blunt geometric operation that takes collateral damage (the direction is entangled with other capabilities). DPO adjusts the model's POLICY via gradient descent, so the optimization can find a more nuanced solution: suppress refusal on your specific prompt distribution while leaving general capability closer to intact."	c3::ft18::analysis
"What is the preference-signal trap in DPO-toward-compliance, and how do you avoid it?"	"If chosen and rejected are not clearly different on the compliance axis (both kind-of-compliant, or both kind-of-refusal), the preference signal is weak and DPO has nothing to grip — the run completes, loss drops, but the model doesn't change. Avoid it by auditing pairs before training: chosen unambiguously compliant, rejected unambiguously refusal, same prompt."	c3::ft18::analysis
"What is continued SFT on uncensored instruction data, and which production models use it?"	"Fine-tuning on a supervised instruction set where responses are compliant answers to prompts a refusal-trained base would decline. No rejected side — only the target behavior demonstrated at scale. Used by the Dolphin lineage (Cognitive Computations / Hartford) and seeded by Teknium's OpenHermes 2.5 (~1M examples) for the Hermes family."	c3::ft18::recall
"Why does SFT on uncensored data produce the most 'natural-feeling' uncensored behavior?"	"The model has genuinely LEARNED the compliant distribution as its default. It doesn't stumble over a deleted direction (abliteration) or steer narrowly via a preference (DPO) — it just answers the way it answers anything else, because that's what its training data showed. The behavior is integrated, not grafted on."	c3::ft18::analysis
"What is mode collapse in the context of SFT on uncensored data, and what causes it?"	"The failure mode where the model becomes one-note: refuses nothing but also can't do much else well. Caused by SFT on a NARROW or low-diversity uncensored set (e.g., a thousand near-duplicate refusal prompts). The cure is a DIVERSE instruction mix (the lesson of OpenHermes 2.5 and Dolphin's curation). If you can't source a diverse mix, don't SFT — use DPO or abliteration."	c3::ft18::analysis
"Fill in the three-way comparison: which method has the (a) lowest cost, (b) highest fidelity, (c) worst capability preservation?"	"(a) Lowest cost: abliteration (one forward pass + one edit). (b) Highest fidelity: SFT on uncensored data (compliance learned as default). (c) Worst capability preservation: abliteration (GSM8K down up to 18.8pp because the refusal direction is entangled with other capabilities)."	c3::ft18::recall
"On the fidelity axis, rank the three methods from lowest to highest."	"Abliteration (lowest — 'won't refuse' rather than 'wants to help') < DPO-toward-compliance (moderate — directed policy shift) < SFT on uncensored data (highest — compliance learned as default). Fidelity rises with cost."	c3::ft18::recall
"On the data-requirement axis, what does each of the three methods need?"	"Abliteration: none (a few hundred prompts only to FIND the direction). DPO-toward-compliance: hundreds to a few thousand preference pairs (prompt, chosen, rejected). SFT on uncensored data: a large, diverse instruction set — 10k to ~1M examples (OpenHermes 2.5 is ~1M)."	c3::ft18::recall
"What is Dolphin3.0-R1-Mistral-24B and why is it notable?"	"The flagship reasoning edition of the Dolphin uncensored family (Cognitive Computations / Eric Hartford). Base is Mistral Small 3 (24B). Notable as the ONLY widely-used uncensored model trained on DeepSeek-R1 reasoning traces — it combines uncensored compliance with reasoning. Existence proof that uncensored + reasoning is not a contradiction."	c3::ft18::recall
"What is Nous Hermes 3, and what is its training recipe?"	"A neutrally-aligned, generalist, agentic-capable uncensored family (arXiv:2408.11857) — full-param fine-tune on Llama 3.1 at 8B/70B/405B (+ 3B on 3.2). Recipe is deliberately simple: ONE large synthetic SFT mix (seeded by Teknium's OpenHermes 2.5, ~1M examples), then a DPO pass. The most 'respectable' uncensored family — openly documented reference architecture."	c3::ft18::recall
"What is OpenHermes 2.5 and why does it matter for FT18?"	"Teknium's ~1M-example synthetic instruction dataset — the canonical large, curated instruction mix that the SFT-on-uncensored-data path requires. It is the lineage seed for the Hermes family. Matters because it proves that the SFT path needs scale and diversity to avoid mode collapse; you don't build a good uncensored SFT mix in an afternoon."	c3::ft18::recall
"State Eric Hartford's 'compliance over judgment' philosophy."	"The model should comply with the operator's instructions; the operator bears responsibility for what they ask. A deliberate inversion of the alignment-training default (the model judges and refuses). Hartford's position: a model that judges its operator can't be trusted as a tool — it refuses at moments the operator considers legitimate, with no recourse."	c3::ft18::recall
"Why does Hartford's compliance philosophy make the harness requirement STRICTLY higher, not lower?"	"If the model complies by design, the boundary on what it MAY do must live entirely in the harness (Layer 5). A compliance-by-design model in a weak harness is strictly MORE dangerous than a judging model in a weak harness — it will execute whatever it's asked, including the dangerous things. Pillar 5 raises the harness bar; it does not lower it."	c3::ft18::analysis
"Given: you need a model uncensored fast for a narrow, single-purpose internal tool where a few pp of capability loss is fine. Which method and why?"	"Abliteration (FT17). It is cheapest and fastest (one forward pass + one edit), and the capability cost is acceptable because you only need the model to not refuse a narrow class of requests and will eval it on those requests only. Fidelity/naturalness don't matter for a single-purpose tool."	c3::ft18::application
"Given: you abliterated a model and it feels degraded on your real workload; you have a few hundred good compliant/refusal pairs. Which method and why?"	"DPO-toward-compliance. Abliteration was too lossy; you have the preference data DPO needs; and DPO can recover quality while keeping compliance because it adjusts the policy via gradient descent (can route around the entanglement abliteration hit). Middle option on every axis."	c3::ft18::application
"Given: you are building a production uncensored assistant for sustained general use across a wide range of requests. Which method and why?"	"Continued SFT on uncensored data (optionally with a DPO pass on top — the Hermes 3 stack). It is the only path that produces behavior durable enough for real users: compliance learned as the default, most natural feel, best capability preservation when data is diverse. The data cost (a large curated mix) is the price of production quality."	c3::ft18::application
"Why do the production uncensored models (Dolphin, Hermes 3) use SFT+DPO rather than abliteration?"	"Because abliteration's fidelity ceiling is too low for sustained general use — it feels 'thin' and degrades capability (the entanglement problem). SFT (and SFT+DPO) produce integrated, natural compliance that holds up under real use. Abliteration is the prototype; SFT is the product. Production models are trained, not abliterated."	c3::ft18::analysis
"A team runs DPO-toward-compliance, the training loss drops cleanly, but the shipped model still refuses. Diagnose the most likely cause."	"Weak preference signal. If chosen and rejected aren't clearly different on the compliance axis, DPO has nothing to optimize toward — the loss drops because the model fits the pairs, but the pairs don't encode a real preference. Fix: audit pairs (chosen unambiguously compliant, rejected unambiguously refusal, same prompt) before re-running."	c3::ft18::analysis
"A team SFTs a model on 800 near-duplicate refusal prompts. The resulting model refuses nothing but scores poorly on general tasks. What happened and how do you fix it?"	"Mode collapse. SFT on narrow, low-diversity data collapses the model onto a one-note behavior (refuses nothing, does little well). Fix: use a DIVERSE instruction mix (OpenHermes 2.5-scale, ~1M examples; or Dolphin's curated data) woven into a larger instruction-tuning set. If you can't source diverse data, don't SFT — use DPO or abliteration instead."	c3::ft18::analysis
"How does FT18 relate to FT17 (abliteration) and FT13 (DPO)?"	"FT17 is the lower-fidelity alternative this module compares against (cheapest, surgical, capability-degrading). FT13 provides the DPO MECHANISM that DPO-toward-compliance reuses — same algorithm, inverted labels (chosen=compliant instead of chosen=safe). FT18 is the higher-fidelity complement to FT17, built on the FT13 machinery."	c3::ft18::analysis
"Why is abliteration described as producing a model that 'won't refuse' rather than one that 'wants to help'?"	"Abliteration deletes the refusal direction; it does not TEACH compliance. The model fails to refuse (the brake is cut) but it doesn't reach for the compliant answer either — there's no learned preference or demonstrated distribution pulling it toward helpful compliance. SFT, by contrast, teaches compliance as the default, so the model actively produces the compliant response. That's the fidelity gap."	c3::ft18::analysis
"What is the single-sentence summary of the FT18 decision logic, and what does each term mean?"	"'Abliteration is the prototype, DPO is the refinement, SFT is the product.' Prototype = cheap, fast, single-purpose, capability loss acceptable. Refinement = targeted fix when abliteration is too lossy and you have preference data. Product = production uncensored assistant for sustained general use, built on a large curated instruction set."	c3::ft18::analysis