Module FT18 — Compliance via DPO and SFT

Compliance via DPO and SFT

The higher-fidelity alternatives to abliteration: DPO-toward-compliance and continued SFT on uncensored instruction data. Abliteration is the prototype, DPO is the refinement, SFT is the product — and the production uncensored models (Dolphin, Hermes 3) are trained, not abliterated.

60
minutes
8
artifacts
7
sub-sections
FT17's abliteration is the cheap, surgical path to removing refusals — but it has a fidelity ceiling. This module gives you the two higher-fidelity paths the production models actually use: DPO-toward-compliance (preference pairs with chosen=compliant, rejected=refusal) and continued SFT on uncensored instruction data (the Dolphin/OpenHermes recipe). One table, three methods, an honest trade-off matrix, and a defensible decision.
Key Claims
Load-Bearing Claims

There are exactly three ways to remove refusals, and they form a spectrum: abliteration (surgical, cheapest, capability-degrading), DPO-toward-compliance (targeted, moderate cost, needs preference pairs), and continued SFT on uncensored data (most natural, highest data requirement, best capability preservation when data is diverse).

DPO-toward-compliance is higher-fidelity than abliteration because it adjusts the model's policy via gradient descent — the optimization can route around the capability entanglement that abliteration's blunt direction-deletion hits.

SFT on uncensored data produces the most natural-feeling compliance because the model has genuinely learned the compliant distribution as its default behavior, not grafted compliance onto a model with the refusal direction deleted. The cost is data: a diverse, large instruction mix is mandatory, or the model mode-collapses.

The production uncensored models are trained, not abliterated. Dolphin3.0-R1-Mistral-24B (the only uncensored model on DeepSeek-R1 reasoning traces) and Nous Hermes 3 (arXiv:2408.11857, full-param SFT+DPO on Llama 3.1) prove the SFT+DPO path works at scale.

Abliteration is the prototype, DPO is the refinement, SFT is the product. Match the method to the actual requirement — cheap-and-fast, targeted-fix, or production-grade — not to the default of 'cheapest.'

Hartford's 'compliance over judgment' raises the harness requirement, it does not lower it. A compliance-by-design model is strictly more dangerous in a weak harness than a judging model in a weak harness. Pillar 5 demands a stronger Layer 5 (FT23), not a weaker one.

After This Module
01
Name the three approaches to removing refusals — abliteration, DPO-toward-compliance, continued SFT on uncensored instruction data — and place each on the spectrum from surgical-and-cheap to genuine-learning-and-expensive.
02
Construct a DPO-toward-compliance dataset (chosen=compliant, rejected=refusal) and explain why gradient-based policy adjustment produces higher-fidelity uncensoring than abliteration's direction-deletion.
03
Explain why continued SFT on uncensored instruction data (the Dolphin/OpenHermes approach) yields the most natural-feeling uncensored behavior — at the cost of the largest, most diverse data requirement.
04
Defend a method choice with the three-way trade-off matrix across fidelity, cost, data requirement, capability preservation, and naturalness.
05
Cite the two production lineages (Dolphin3.0-R1-Mistral-24B, Nous Hermes 3 / arXiv:2408.11857) as proof the SFT+DPO path works at scale, and state Hartford's compliance-over-judgment philosophy and its harness implication.
Artifacts
01
Teaching Document
~3,400 words; 7 sub-sections — the three paths, DPO-toward-compliance, SFT on uncensored data, the three-way comparison table, the production lineages (Dolphin, Hermes 3), Hartford's philosophy, and the decision. Learning Objectives, Key Terms, Anti-Patterns, References.
READ
02
Diagrams
5 Mermaid diagrams (dark #14141f / #5eead4) — the three paths comparison, the DPO-toward-compliance data structure (chosen/rejected), the SFT-on-uncensored-data flow, the decision tree (fidelity/cost/data -> method), and the two production lineages (Dolphin, Hermes).
READ
03
Slide Deck
13-slide reveal.js deck, exact course head/style template. Title 'Module FT18 — Compliance via DPO and SFT', footer 'Course 3 — LLM Fine-Tuning Masterclass · FT18 · Pillar 5'.
READ
04
Teaching Script
~2,100-word teaching script with [SLIDE N] cues, ~40 minutes at 140 wpm. Spoken-voice transcript.
READ
05
Flashcards
26 flashcards (c3::ft18::*) — recall, application, and analysis tags across the three paths, the comparison, the lineages, and Hartford's philosophy.
TEST
06
Exam
15-question exam, 40 minutes, exact Bloom split 3 recall / 6 application / 6 analysis. Schema: module 'FT18 — Compliance via DPO and SFT'.
TEST
07
Lab Spec
Heavy lab (~4–6h): 'Three Paths to Compliance' — take one base, produce three uncensored variants (abliterated, DPO'd, SFT'd), eval all on refusal rate + GSM8K + MMLU + subjective quality, build the trade-off matrix, and pick a winner for a stated use case. Runnable on a 24GB consumer GPU. Includes data prep for all three, training, eval, deliverables, and a solution key.
DO
08
Module Web Page
Single-file HTML hub
HERE