Module FTDD-06 — Dolphin / Hermes — Uncensored Lineages as Engineering Case Studies

Dolphin / Hermes — Uncensored Lineages as Engineering Case Studies

The uncensored model lineages — Eric Hartford's Dolphin series and Nous Research's Hermes 3 — studied as engineering case studies, not advocacy. They are production examples of the FT16–FT18 alignment-control techniques: dataset curation for compliance-over-judgment, full-param SFT+DPO on Llama 3.1, and reasoning-trace training on DeepSeek-R1. The lesson is the recipe and the trade-offs, not the ideology.

45
minutes
8
artifacts
4
sub-sections
Dolphin3.0-R1-Mistral-24B is the only uncensored model trained on DeepSeek-R1 reasoning traces. Hermes 3 is a full-parameter SFT+DPO steer of Llama 3.1 at 8B, 70B, and 405B, described in arXiv:2408.11857 as 'unlocked, uncensored, highly steerable.' Behind both is the OpenHermes 2.5 dataset — roughly one million examples curated by Teknium. Study these as engineering: what the recipe is, what it costs, and why the model only becomes responsible inside a harness.
Key Claims
Load-Bearing Claims

These are engineering case studies, not advocacy. Dolphin and Hermes are studied for their recipes (dataset curation, full-param SFT+DPO, reasoning-trace RL) and their trade-offs (capability cost of uncensoring). The course's stance — the model steers, the harness bounds (FT00) — is unchanged: an uncensored model is only responsible inside an eval'd harness.

Hermes 3 (Nous Research, arXiv:2408.11857) is the canonical full-param alignment-control recipe. SFT then DPO on Llama 3.1 at 8B/70B/405B, on primarily synthetic data, producing a 'neutrally-aligned, highly steerable' model. It is a production example of the FT12 (SFT) + FT13 (DPO) + FT16–18 (alignment control) stack at scale.

Dolphin (Eric Hartford, Cognitive Computations) is the compliance-over-judgment philosophy, operationalized. Dolphin3.0-R1-Mistral-24B is the only uncensored model trained on DeepSeek-R1 reasoning traces — it pairs Hartford's refusal-removal philosophy with R1-grade reasoning via FT14-style verifiable-reward RL on reasoning traces.

OpenHermes 2.5 (Teknium, ~1M examples) is the shared dataset backbone. It trained both the OpenHermes 2.5 and Nous Hermes 2 model families and is the textbook example of the course's thesis: the dataset is the steering wheel (FT00). Study the lineage to see how a single high-quality dataset propagates through a family of steered models.

After This Module
01
Describe the Hermes 3 recipe (full-param SFT then DPO on Llama 3.1, primarily synthetic data) and place each stage on the Steering Stack (FT12 SFT, FT13 DPO, FT16–18 alignment control).
02
Describe the Dolphin lineage and philosophy (Eric Hartford's compliance-over-judgment) and explain what makes Dolphin3.0-R1-Mistral-24B technically distinctive (R1 reasoning traces on an uncensored base).
03
Explain the role of OpenHermes 2.5 (~1M examples) as the dataset backbone of the Hermes family, and connect it to the course thesis that the dataset is the steering wheel.
04
State, as engineering trade-offs rather than ideology, the capability cost of uncensoring and why these models are only responsible inside an eval'd harness (the FT00 / FT23 synthesis).
Artifacts