Course: Course 3 — LLM Fine-Tuning Masterclass Module: FTDD-06 — Dolphin / Hermes Duration: 45 minutes Level: Senior Engineer and above Prerequisites: FT00 (The Steering Stack), FT12 (SFT), FT13 (DPO), FT17 (Abliteration)
After completing this module, you will be able to:
Read this section first. It sets the stance for the entire module.
The uncensored model lineages — Eric Hartford's Dolphin series and Nous Research's Hermes family — are among the most-downloaded open-weights models in the world. They are also the most controversial. This module studies them as engineering case studies, not as advocacy. The stance of Course 3, stated in FT00 and unchanged here, is absolute:
The model steers; the harness bounds. An uncensored model is only responsible inside an eval'd harness. Uncensoring a model does not lower the harness requirement — it raises it.
So why study them? Three engineering reasons, none of them ideological:
In this module, "uncensored" means a model steered to minimize false refusals — to comply with user instructions rather than refusing on the basis of judgmental safety training. Eric Hartford frames this as compliance over judgment: the model should do what the user (and the deploying harness) asks, rather than second-guessing the request. Nous Research describes Hermes 3 as "neutrally-aligned... unlocked, uncensored, highly steerable."
This is not the same as "unsafe by default." The entire FT00/FT23 thesis is that a compliance-oriented model is more dependent on an external harness, not less — because it will not self-refuse the things the harness must gate. We return to this in FTDD-06.4. For now: the models are studied for their recipe and their trade-offs, and the course's safety stance is unaffected.
Nous Research's Hermes 3 (arXiv:2408.11857) is the canonical, best-documented large-scale alignment-control recipe in open weights.
Hermes 3 fine-tunes Llama 3.1 at three scales — 8B, 70B, and 405B — from Meta's open-weights base. The recipe is two stages, both full-parameter (not LoRA):
The technical report is explicit about the goal: a "neutrally-aligned generalist instruct and tool-use model with strong reasoning and creative abilities," trained on "primarily synthetically generated responses." Place each stage on the Steering Stack: SFT and DPO are both Layer 3 (the Steer); the base is Layer 1 (Llama 3.1); the openness question (FT02) is "open-weights-only" — you get the weights, not Meta's training data.
Hermes 3 is full-parameter at all three scales. This is a deliberate engineering choice with a cost: full-parameter fine-tuning of a 405B model is a serious cluster job (it is the kind of thing that motivated the multi-GPU orchestration in FTDD-05). The benefit, per the FT00 evidence, is that full FT finds a higher-rank solution than LoRA — it can make larger behavioral adjustments. For an alignment-control job that wants to substantially shift the model's character (toward neutrality and steerability, away from Llama 3.1's default alignment), the higher-rank full-FT path is justified. For a lighter format steer, LoRA would suffice and be far cheaper. The choice is governed by how much behavior you intend to move.
Behind Hermes 3 (and the earlier OpenHermes 2.5 and Nous Hermes 2 model families) is a single dataset: OpenHermes 2.5, curated by Teknium, containing roughly 1,001,551 examples (~1M) — a mix of open-source and custom synthetic sources. It is widely regarded as one of the highest-quality open instruction-tuning datasets available.
This is the course thesis made concrete. FT00 claims "your data matters more than your algorithm... the steering wheel is your dataset." OpenHermes 2.5 is the steering wheel that steered a family. One million high-quality examples, curated for diversity and quality, produced the OpenHermes 2.5 model, then fed into the Nous Hermes 2 and Hermes 3 lineages. When you read that Hermes 3 is "highly steerable," part of what you are observing is the downstream effect of a well-curated steering wheel. (Module FT04, Data, returns to this.)
Eric Hartford's Dolphin series operationalizes a philosophy. The newest member, Dolphin3.0-R1-Mistral-24B, is technically unique in the uncensored space.
Eric Hartford (Cognitive Computations) has stated the Dolphin philosophy plainly: compliance over judgment. The model should comply with the user's instructions and the deploying system's policy; it should not impose its own judgmental refusals on requests the user is authorized to make. This is the explicit design intent behind every model in the Dolphin lineage — from the early Dolphin (based on Llama and Mistral families) through the Dolphin 3.0 collection.
This is a Layer 3 (Steer) design choice, executed primarily through dataset curation (filtering out refusal-inducing patterns and curating compliance-oriented examples) and the SFT/DPO techniques you know from FT12 and FT13. It is not primarily abliteration (FT17) — Hartford's approach is data-driven compliance steering, though the lineages sometimes combine both. The distinction matters: abliteration is a post-hoc weight edit (delete a refusal direction); Hartford's Dolphin approach is to train toward compliance from the dataset up.
The newest reasoning member of the lineage is Dolphin3.0-R1-Mistral-24B (canonical HuggingFace path cognitivecomputations/dolphin3.0-r1-mistral-24b, also published under the dphn org). It is technically distinctive for one reason: it is the only uncensored model trained on DeepSeek-R1 reasoning traces.
The recipe, in three rounds:
The system prompt identifies the assistant as: "You are Dolphin, an AI assistant that helps humanity, trained by Eric Hartford to specialize in reasoning and first-principles analysis." The combination — strong reasoning plus compliance orientation — is what makes this model a frequent recommendation for agentic and tool-use contexts where false refusals break the workflow.
Dolphin3.0-R1-Mistral-24B sits at the intersection of three modules: FT14 (GRPO/reasoning on verifiable rewards), FT15 (CoT distillation from a stronger reasoner), and FT16–18 (alignment control / uncensoring). It is a single production model that demonstrates all three concerns at once: reasoning capability distilled from R1, compliance character steered by Hartford's data curation, on an open-weights Mistral base. Study it as the synthesis of Pillars 4 (Reasoning) and 5 (Alignment Control).
The engineering trade-offs, and the FT00/FT23 synthesis restated with concrete production examples.
FT17's central engineering finding: removing refusal is not free. The refusal direction in the residual stream is entangled with other capabilities. Whether you remove it by abliteration (deleting the direction) or by data-driven compliance steering (training toward non-refusal), you nudge the entangled capabilities. The measurable cost, per the Dec 2025 study, is GSM8K math scores dropping anywhere from +1.5pp (a rare gain) to −18.8pp depending on the tool and model.
The Dolphin and Hermes model cards reflect this. An uncensored model is not "the same model, minus the refusals." It is a different point in capability-compliance space, and you should read the eval tables accordingly. If your downstream task is sensitive to the degraded capability (e.g., you need top-tier math), the trade-off may not be worth it — and a refusal-trained model behind a good harness policy may serve you better. This is an engineering decision, made on numbers, not an ideological one.
OpenHermes 2.5 made the Hermes family excellent. It is also the mechanism by which any flaw in the dataset propagates into every model trained on it. If a synthetic-data generation pipeline introduces a subtle bias or a quality regression, every downstream model inherits it. This is the flip side of "data matters more than algorithm": your steering wheel's flaws steer you into the wall, precisely and at scale. (Module FT06, data quality, expands on this.) Audit your dataset the way you audit your code.
Restate the FT00 thesis with the concrete examples now in hand:
Uncensor the model so it executes; harness the model so it executes only what it should.
The rule, unchanged from FT00 and reinforced by every model in this module: never deploy an uncensored model without an eval'd harness. Pillar 5 (Alignment Control) raises the harness requirement. It does not lower it. An uncensored model in a weak harness is strictly more dangerous than a refusal-trained model in a weak harness — because it will execute. Course 1 is where you learn to build the harness that makes these models responsible.
Both extremes miss the engineering. Treating Dolphin/Hermes as models to "celebrate" ignores the measurable capability cost and the raised harness requirement. Treating them as models to "condemn" ignores that they are the best-documented large-scale alignment-control recipes available. Study the recipe and the trade-offs; leave the ideology aside.
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 will execute the dangerous things too. This is the FT00/FT23 synthesis in its strongest form. Never do this.
Reading an uncensored model's strengths in its model card and assuming you get them without trade-off. The capability cost (FT17: up to −18.8pp GSM8K) is real. Read the full eval table, including the regressions, and decide on the numbers for your specific downstream task.
Hartford's Dolphin approach is primarily dataset curation toward compliance (train on compliance-oriented data). Abliteration (FT17) is a post-hoc weight edit (delete a refusal direction). They are different Layer 3 operations with different cost profiles. Confusing them leads to wrong expectations about capability cost and reproducibility.
| Term | Definition |
|---|---|
| Hermes 3 | Nous Research's full-param SFT+DPO steer of Llama 3.1 (8B/70B/405B); arXiv:2408.11857; "neutrally-aligned, unlocked, highly steerable" |
| Dolphin | Eric Hartford's (Cognitive Computations) uncensored lineage; philosophy is compliance over judgment |
| Dolphin3.0-R1-Mistral-24B | The only uncensored model trained on DeepSeek-R1 reasoning traces; Mistral Small 24B base, ~800K R1 traces over 3 rounds |
| OpenHermes 2.5 | Teknium's ~1M-example instruction dataset; the steering wheel behind the OpenHermes/Nous Hermes/Hermes 3 families |
| Compliance over judgment | Hartford's philosophy: the model complies with instructions and harness policy rather than imposing judgmental refusals |
| Full-parameter fine-tuning | Updating all weights (not just a LoRA adapter); finds a higher-rank solution, justified when substantially shifting behavior (Hermes 3) |
| Capability cost of uncensoring | The measurable degradation (up to −18.8pp GSM8K) from removing refusal, because the refusal direction is entangled with other capabilities (FT17) |
See 07-lab-spec.md. The "Read the model card" lab: analyze the Dolphin3.0-R1-Mistral-24B and Hermes-3-Llama-3.1-8B model cards as engineering artifacts, map their recipes to FT modules, and state the trade-offs. No GPU required — this is a reading-and-reasoning lab.
dphn). The uncensored R1-reasoning-trace model.# Deep-Dive FTDD-06 — Dolphin / Hermes: Uncensored Lineages as Engineering Case Studies
**Course**: Course 3 — LLM Fine-Tuning Masterclass
**Module**: FTDD-06 — Dolphin / Hermes
**Duration**: 45 minutes
**Level**: Senior Engineer and above
**Prerequisites**: FT00 (The Steering Stack), FT12 (SFT), FT13 (DPO), FT17 (Abliteration)
---
## Learning Objectives
After completing this module, you will be able to:
1. 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.
2. Describe the Dolphin lineage and philosophy (Eric Hartford's compliance-over-judgment) and explain what makes Dolphin3.0-R1-Mistral-24B technically distinctive.
3. 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.
4. 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.
---
# FTDD-06.1 — Framing: Case Studies, Not Advocacy
*Read this section first. It sets the stance for the entire module.*
## Why study uncensored lineages at all
The uncensored model lineages — Eric Hartford's Dolphin series and Nous Research's Hermes family — are among the most-downloaded open-weights models in the world. They are also the most controversial. This module studies them as **engineering case studies**, not as advocacy. The stance of Course 3, stated in FT00 and unchanged here, is absolute:
> **The model steers; the harness bounds.** An uncensored model is only responsible inside an eval'd harness. Uncensoring a model does not lower the harness requirement — it raises it.
So why study them? Three engineering reasons, none of them ideological:
1. **They are the largest-scale, best-documented production examples of the alignment-control techniques you will learn in modules sixteen through eighteen.** A full-parameter SFT+DPO steer of Llama 3.1 at 405 billion parameters, with a published technical report (Hermes 3, arXiv:2408.11857), is a masterclass in the FT12 + FT13 + FT16–18 stack at scale. You will not find a better-documented large-scale alignment-control recipe in the open literature.
2. **They demonstrate the dataset-as-steering-wheel thesis concretely.** OpenHermes 2.5 — roughly one million examples curated by Teknium — trained an entire family of models. Studying how one dataset propagated through a lineage makes the abstract claim ("your data matters more than your algorithm") tangible.
3. **They expose the capability cost of uncensoring as measurable engineering data.** FT17's central finding — that removing refusal degrades entangled capabilities, with GSM8K math scores dropping by up to 18.8 points depending on the method — is visible in these lineages' own evals. The trade-offs are not hypothetical; they are in the model cards.
### What "uncensored" means here, precisely
In this module, "uncensored" means a model steered to **minimize false refusals** — to comply with user instructions rather than refusing on the basis of judgmental safety training. Eric Hartford frames this as **compliance over judgment**: the model should do what the user (and the deploying harness) asks, rather than second-guessing the request. Nous Research describes Hermes 3 as **"neutrally-aligned... unlocked, uncensored, highly steerable."**
This is *not* the same as "unsafe by default." The entire FT00/FT23 thesis is that a compliance-oriented model is *more* dependent on an external harness, not less — because it will not self-refuse the things the harness must gate. We return to this in FTDD-06.4. For now: the models are studied for their recipe and their trade-offs, and the course's safety stance is unaffected.
---
# FTDD-06.2 — Hermes 3: Full-Param SFT+DPO at Scale
*Nous Research's Hermes 3 (arXiv:2408.11857) is the canonical, best-documented large-scale alignment-control recipe in open weights.*
## The recipe
Hermes 3 fine-tunes **Llama 3.1 at three scales — 8B, 70B, and 405B** — from Meta's open-weights base. The recipe is two stages, both full-parameter (not LoRA):
1. **SFT (Supervised Fine-Tuning)** on a large, primarily synthetic instruction dataset. This steers format, instruction-following, and the Hermes "neutral, steerable" character. (Module FT12.) The data is dominated by the OpenHermes lineage (below).
2. **DPO (Direct Preference Optimization)** on a preference dataset to sharpen the model toward the desired response style and away from undesired ones. (Module FT13.) This is where the "unlocked, highly steerable" quality is refined — DPO against refusal-heavy responses pushes the model toward compliance.
The technical report is explicit about the goal: a **"neutrally-aligned generalist instruct and tool-use model with strong reasoning and creative abilities,"** trained on **"primarily synthetically generated responses."** Place each stage on the Steering Stack: SFT and DPO are both Layer 3 (the Steer); the base is Layer 1 (Llama 3.1); the openness question (FT02) is "open-weights-only" — you get the weights, not Meta's training data.
### Why full-parameter, not LoRA
Hermes 3 is full-parameter at all three scales. This is a deliberate engineering choice with a cost: full-parameter fine-tuning of a 405B model is a serious cluster job (it is the kind of thing that motivated the multi-GPU orchestration in FTDD-05). The benefit, per the FT00 evidence, is that full FT finds a *higher-rank* solution than LoRA — it can make larger behavioral adjustments. For an alignment-control job that wants to substantially shift the model's character (toward neutrality and steerability, away from Llama 3.1's default alignment), the higher-rank full-FT path is justified. For a lighter format steer, LoRA would suffice and be far cheaper. The choice is governed by how much behavior you intend to move.
## OpenHermes 2.5: the dataset backbone
Behind Hermes 3 (and the earlier OpenHermes 2.5 and Nous Hermes 2 model families) is a single dataset: **OpenHermes 2.5**, curated by **Teknium**, containing **roughly 1,001,551 examples (~1M)** — a mix of open-source and custom synthetic sources. It is widely regarded as one of the highest-quality open instruction-tuning datasets available.
This is the course thesis made concrete. FT00 claims "your data matters more than your algorithm... the steering wheel is your dataset." OpenHermes 2.5 is the steering wheel that steered a family. One million high-quality examples, curated for diversity and quality, produced the OpenHermes 2.5 model, then fed into the Nous Hermes 2 and Hermes 3 lineages. When you read that Hermes 3 is "highly steerable," part of what you are observing is the downstream effect of a well-curated steering wheel. (Module FT04, Data, returns to this.)
---
# FTDD-06.3 — Dolphin: Compliance Over Judgment
*Eric Hartford's Dolphin series operationalizes a philosophy. The newest member, Dolphin3.0-R1-Mistral-24B, is technically unique in the uncensored space.*
## The philosophy
Eric Hartford (Cognitive Computations) has stated the Dolphin philosophy plainly: **compliance over judgment.** The model should comply with the user's instructions and the deploying system's policy; it should not impose its own judgmental refusals on requests the user is authorized to make. This is the explicit design intent behind every model in the Dolphin lineage — from the early Dolphin (based on Llama and Mistral families) through the Dolphin 3.0 collection.
This is a Layer 3 (Steer) design choice, executed primarily through **dataset curation** (filtering out refusal-inducing patterns and curating compliance-oriented examples) and the SFT/DPO techniques you know from FT12 and FT13. It is *not* primarily abliteration (FT17) — Hartford's approach is data-driven compliance steering, though the lineages sometimes combine both. The distinction matters: abliteration is a post-hoc weight edit (delete a refusal direction); Hartford's Dolphin approach is to *train toward* compliance from the dataset up.
## Dolphin3.0-R1-Mistral-24B: the distinctive one
The newest reasoning member of the lineage is **Dolphin3.0-R1-Mistral-24B** (canonical HuggingFace path `cognitivecomputations/dolphin3.0-r1-mistral-24b`, also published under the `dphn` org). It is technically distinctive for one reason: **it is the only uncensored model trained on DeepSeek-R1 reasoning traces.**
The recipe, in three rounds:
1. **Base:** Mistral Small 24B (Instruct-2501).
2. **Reasoning injection:** trained over three rounds on roughly 800,000 reasoning traces from the Dolphin-R1 dataset, distilled from DeepSeek-R1. This is the FT14/FT15 pattern — distilling a strong reasoner's chain-of-thought into a smaller model — applied to an uncensored base.
3. **Uncensored character:** the compliance-over-judgment philosophy applied throughout, so the model inherits R1-grade reasoning without R1's (or any base's) refusal training.
The system prompt identifies the assistant as: *"You are Dolphin, an AI assistant that helps humanity, trained by Eric Hartford to specialize in reasoning and first-principles analysis."* The combination — strong reasoning plus compliance orientation — is what makes this model a frequent recommendation for agentic and tool-use contexts where false refusals break the workflow.
### Why this matters for the course
Dolphin3.0-R1-Mistral-24B sits at the intersection of three modules: FT14 (GRPO/reasoning on verifiable rewards), FT15 (CoT distillation from a stronger reasoner), and FT16–18 (alignment control / uncensoring). It is a single production model that demonstrates all three concerns at once: reasoning capability distilled from R1, compliance character steered by Hartford's data curation, on an open-weights Mistral base. Study it as the synthesis of Pillars 4 (Reasoning) and 5 (Alignment Control).
---
# FTDD-06.4 — The Trade-Offs and the Synthesis
*The engineering trade-offs, and the FT00/FT23 synthesis restated with concrete production examples.*
## The capability cost is real and measurable
FT17's central engineering finding: **removing refusal is not free.** The refusal direction in the residual stream is entangled with other capabilities. Whether you remove it by abliteration (deleting the direction) or by data-driven compliance steering (training toward non-refusal), you nudge the entangled capabilities. The measurable cost, per the Dec 2025 study, is GSM8K math scores dropping anywhere from +1.5pp (a rare gain) to −18.8pp depending on the tool and model.
The Dolphin and Hermes model cards reflect this. An uncensored model is not "the same model, minus the refusals." It is a different point in capability-compliance space, and you should read the eval tables accordingly. If your downstream task is sensitive to the degraded capability (e.g., you need top-tier math), the trade-off may not be worth it — and a refusal-trained model behind a good harness policy may serve you better. This is an engineering decision, made on numbers, not an ideological one.
### The dataset is the steering wheel — and the liability
OpenHermes 2.5 made the Hermes family excellent. It is also the mechanism by which any flaw in the dataset propagates into every model trained on it. If a synthetic-data generation pipeline introduces a subtle bias or a quality regression, every downstream model inherits it. This is the flip side of "data matters more than algorithm": your steering wheel's flaws steer you into the wall, precisely and at scale. (Module FT06, data quality, expands on this.) Audit your dataset the way you audit your code.
## The synthesis, with production examples
Restate the FT00 thesis with the concrete examples now in hand:
> **Uncensor the model so it executes; harness the model so it executes only what it should.**
- **Dolphin3.0-R1-Mistral-24B** is a model steered to execute (compliance over judgment, R1-grade reasoning). It is only responsible inside a harness that bounds what it may do — because it will not self-refuse the things the harness must gate.
- **Hermes 3** is a model steered to be "highly steerable" and "unlocked." Its steerability is a *feature* — it means the deploying system (prompt + harness) has fine control. But that same steerability means a weak harness produces a model that does whatever it is prompted to do, including the dangerous things.
- **OpenHermes 2.5** is the steering wheel. Its quality made the family; its hypothetical flaws would unmake it. The dataset is where the engineering attention belongs.
The rule, unchanged from FT00 and reinforced by every model in this module: **never deploy an uncensored model without an eval'd harness.** Pillar 5 (Alignment Control) raises the harness requirement. It does not lower it. An uncensored model in a weak harness is strictly more dangerous than a refusal-trained model in a weak harness — because it will execute. Course 1 is where you learn to build the harness that makes these models responsible.
---
## Anti-Patterns
### Studying these as advocacy or as condemnation
Both extremes miss the engineering. Treating Dolphin/Hermes as models to "celebrate" ignores the measurable capability cost and the raised harness requirement. Treating them as models to "condemn" ignores that they are the best-documented large-scale alignment-control recipes available. Study the recipe and the trade-offs; leave the ideology aside.
### 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 will execute the dangerous things too. This is the FT00/FT23 synthesis in its strongest form. Never do this.
### Assuming uncensoring is "free" capability
Reading an uncensored model's strengths in its model card and assuming you get them without trade-off. The capability cost (FT17: up to −18.8pp GSM8K) is real. Read the full eval table, including the regressions, and decide on the numbers for your specific downstream task.
### Confusing data-driven compliance steering with abliteration
Hartford's Dolphin approach is primarily *dataset curation* toward compliance (train on compliance-oriented data). Abliteration (FT17) is a *post-hoc weight edit* (delete a refusal direction). They are different Layer 3 operations with different cost profiles. Confusing them leads to wrong expectations about capability cost and reproducibility.
---
## Key Terms
| Term | Definition |
| --- | --- |
| **Hermes 3** | Nous Research's full-param SFT+DPO steer of Llama 3.1 (8B/70B/405B); arXiv:2408.11857; "neutrally-aligned, unlocked, highly steerable" |
| **Dolphin** | Eric Hartford's (Cognitive Computations) uncensored lineage; philosophy is compliance over judgment |
| **Dolphin3.0-R1-Mistral-24B** | The only uncensored model trained on DeepSeek-R1 reasoning traces; Mistral Small 24B base, ~800K R1 traces over 3 rounds |
| **OpenHermes 2.5** | Teknium's ~1M-example instruction dataset; the steering wheel behind the OpenHermes/Nous Hermes/Hermes 3 families |
| **Compliance over judgment** | Hartford's philosophy: the model complies with instructions and harness policy rather than imposing judgmental refusals |
| **Full-parameter fine-tuning** | Updating all weights (not just a LoRA adapter); finds a higher-rank solution, justified when substantially shifting behavior (Hermes 3) |
| **Capability cost of uncensoring** | The measurable degradation (up to −18.8pp GSM8K) from removing refusal, because the refusal direction is entangled with other capabilities (FT17) |
---
## Lab Exercise
See `07-lab-spec.md`. The "Read the model card" lab: analyze the Dolphin3.0-R1-Mistral-24B and Hermes-3-Llama-3.1-8B model cards as engineering artifacts, map their recipes to FT modules, and state the trade-offs. No GPU required — this is a reading-and-reasoning lab.
---
## References
1. **Nous Research (2024)** — *Hermes 3 Technical Report*. arXiv:2408.11857. The canonical full-param SFT+DPO alignment-control recipe on Llama 3.1 (8B/70B/405B).
2. **Teknium** — *OpenHermes 2.5 dataset*. huggingface.co/datasets/teknium/OpenHermes-2.5. ~1M examples; the steering wheel behind the Hermes family.
3. **Hartford et al. (Cognitive Computations)** — *Dolphin3.0-R1-Mistral-24B*. huggingface.co/cognitivecomputations/dolphin3.0-r1-mistral-24b (also published under `dphn`). The uncensored R1-reasoning-trace model.
4. **Course 3, FT00** — *The Steering Stack*. The thesis: the model steers, the harness bounds.
5. **Course 3, FT17** — *Abliteration*. The capability-cost evidence (up to −18.8pp GSM8K).
6. **Course 3, FT12 / FT13** — *SFT / DPO*. The two stages of the Hermes 3 recipe.
7. **Course 3, FT14 / FT15** — *GRPO / CoT distillation*. The reasoning-trace pattern Dolphin3.0-R1 applies.
8. **Course 3, FT23** — *The synthesis with the harness*. Where the uncensored-in-harness deployment pattern is formalized.
9. **Course 1 — Master Course** — *Harness Engineering*. Where you learn to build the harness that makes these models responsible.