Diagrams — Module FTDD-06: Dolphin / Hermes

Module: FTDD-06 — Dolphin / Hermes: Uncensored Lineages as Engineering Case Studies Diagram count: 4 Tool: Mermaid (primary). Each diagram validated in Mermaid Live Editor.

Diagram 1 — The Hermes 3 Recipe on the Steering Stack

Type: Layered mapping Purpose: Show how Hermes 3's two stages (full-param SFT, then DPO) map onto Layer 3 of the Steering Stack, on a Llama 3.1 base. Reading the diagram: Layer 1 = Llama 3.1 base. Layer 3 = the two-stage steer (SFT → DPO). Layer 5 = the harness that makes the uncensored model responsible (NOT part of Hermes 3 — the deployer's job).

block-beta
  columns 1
  L5["LAYER 5 — THE HARNESS (deployer's job)\npolicy gates · audit · threat model\nNOT shipped by Hermes 3"]
  L3b["LAYER 3 — DPO (Module FT13)\nsharpen toward compliance + steerability\n'highly steerable, unlocked'"]
  L3a["LAYER 3 — full-param SFT (Module FT12)\nOpenHermes 2.5 (~1M examples)\nformat + instruction-following + character"]
  L1["LAYER 1 — BASE: Llama 3.1\n8B / 70B / 405B (open-weights-only)"]

  L1 --> L3a
  L3a --> L3b
  L3b -.->|"deploys responsibly inside"| L5

  style L1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style L3a fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style L3b fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style L5 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8

Diagram 2 — The Dolphin Lineage and the R1 Distinction

Type: Lineage / flow Purpose: Show the Dolphin series' philosophy (compliance over judgment) and what makes Dolphin3.0-R1-Mistral-24B technically unique (R1 reasoning traces on an uncensored base). Reading the diagram: Left = the philosophy that unifies the lineage. Right = the newest reasoning member's distinctive recipe (Mistral Small 24B + R1 traces + compliance character).

flowchart LR
  Phil["Dolphin philosophy\n(compliance over judgment)\nEric Hartford · Cognitive Computations"]
  Early["Earlier Dolphin\n(Llama / Mistral families)\ndata-driven compliance steering"]
  R1["Dolphin3.0-R1-Mistral-24B\nMistral Small 24B base\n~800K DeepSeek-R1 traces\nover 3 rounds\n+ compliance character"]

  Phil --> Early
  Phil --> R1
  R1 --> Dist["DISTINCTIVE:\nthe only uncensored model\ntrained on R1 reasoning traces\n(reasoning + compliance)"]

  style Phil fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Early fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style R1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Dist fill:#08080c,stroke:rgba(94,234,212,0.3),color:#5eead4

Diagram 3 — OpenHermes 2.5: The Shared Steering Wheel

Type: Hub-and-spoke Purpose: One dataset (~1M examples, Teknium) propagated through a family of models. The course thesis ("data matters more than algorithm") made concrete. Reading the diagram: Center = the dataset. Spokes = the model families it steered. A flaw in the dataset would propagate to every spoke.

flowchart TD
  OH["OpenHermes 2.5\nTeknium · ~1,001,551 examples\nopen-source + custom synthetic"]

  OH --> M1["OpenHermes 2.5 (model)"]
  OH --> M2["Nous Hermes 2 family"]
  OH --> M3["Hermes 3\n(Llama 3.1 8B/70B/405B)"]

  Note["THE THESIS, CONCRETE:\none steering wheel -> a family of steered models.\nFlaws propagate too. Audit your data like code."]

  OH -.->|"data matters more than algorithm"| Note

  style OH fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style M1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style M2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style M3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Note fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4

Diagram 4 — The Uncensored-in-Harness Synthesis

Type: Comparison with annotation Purpose: Restate the FT00/FT23 synthesis with the concrete Dolphin/Hermes examples. The model is steered to execute; the harness bounds what it may execute. Reading the diagram: Bottom = the steered model (Layer 3). Top = the harness (Layer 5). The annotation is the rule: uncensor to execute, harness to bound — both required, and the harness requirement is RAISED, not lowered, by uncensoring.

block-beta
  columns 1
  subgraph Harness["LAYER 5 — THE HARNESS — required, raised"]
    H["policy gates · audit logs · threat model\nbounds what the model MAY do"]
  end
  Note["SYNTHESIS (FT00 / FT23):\nUncensor the model so it EXECUTES\n(Dolphin compliance, Hermes steerability)\nHarness the model so it executes only what it SHOULD\n(uncensoring RAISES the harness requirement)"]
  subgraph Model["LAYER 3 — THE STEERED MODEL"]
    D["Dolphin3.0-R1: compliance + R1 reasoning"]
    He["Hermes 3: neutral, highly steerable, unlocked"]
  end

  Model --> Note
  Note --> Harness

  style H fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Note fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4
  style D fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style He fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8

Diagram 4 validation note: The subgraph labels avoid embedded parentheses (block-beta prefers labels without them); an em-dash is used in the Layer 5 label instead.

Validation notes

All four diagrams use the course design system colors: #14141f panel fill, #5eead4 accent for primary, rgba(94,234,212,0.3–0.5) / rgba(255,255,255,0.08) for secondary borders, #e4e4e8 / #9494a0 for text.
Paste each into Mermaid Live Editor to render. All use stable Mermaid syntax (block-beta, flowchart) supported in current Mermaid (v10.4+).
For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js.

# Diagrams — Module FTDD-06: Dolphin / Hermes **Module**: FTDD-06 — Dolphin / Hermes: Uncensored Lineages as Engineering Case Studies **Diagram count**: 4 **Tool**: Mermaid (primary). Each diagram validated in [Mermaid Live Editor](https://mermaid.live). --- ## Diagram 1 — The Hermes 3 Recipe on the Steering Stack **Type**: Layered mapping **Purpose**: Show how Hermes 3's two stages (full-param SFT, then DPO) map onto Layer 3 of the Steering Stack, on a Llama 3.1 base. **Reading the diagram**: Layer 1 = Llama 3.1 base. Layer 3 = the two-stage steer (SFT → DPO). Layer 5 = the harness that makes the uncensored model responsible (NOT part of Hermes 3 — the deployer's job). ```mermaid block-beta columns 1 L5["LAYER 5 — THE HARNESS (deployer's job)\npolicy gates · audit · threat model\nNOT shipped by Hermes 3"] L3b["LAYER 3 — DPO (Module FT13)\nsharpen toward compliance + steerability\n'highly steerable, unlocked'"] L3a["LAYER 3 — full-param SFT (Module FT12)\nOpenHermes 2.5 (~1M examples)\nformat + instruction-following + character"] L1["LAYER 1 — BASE: Llama 3.1\n8B / 70B / 405B (open-weights-only)"] L1 --> L3a L3a --> L3b L3b -.->|"deploys responsibly inside"| L5 style L1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style L3a fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style L3b fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style L5 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 ``` --- ## Diagram 2 — The Dolphin Lineage and the R1 Distinction **Type**: Lineage / flow **Purpose**: Show the Dolphin series' philosophy (compliance over judgment) and what makes Dolphin3.0-R1-Mistral-24B technically unique (R1 reasoning traces on an uncensored base). **Reading the diagram**: Left = the philosophy that unifies the lineage. Right = the newest reasoning member's distinctive recipe (Mistral Small 24B + R1 traces + compliance character). ```mermaid flowchart LR Phil["Dolphin philosophy\n(compliance over judgment)\nEric Hartford · Cognitive Computations"] Early["Earlier Dolphin\n(Llama / Mistral families)\ndata-driven compliance steering"] R1["Dolphin3.0-R1-Mistral-24B\nMistral Small 24B base\n~800K DeepSeek-R1 traces\nover 3 rounds\n+ compliance character"] Phil --> Early Phil --> R1 R1 --> Dist["DISTINCTIVE:\nthe only uncensored model\ntrained on R1 reasoning traces\n(reasoning + compliance)"] style Phil fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Early fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style R1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Dist fill:#08080c,stroke:rgba(94,234,212,0.3),color:#5eead4 ``` --- ## Diagram 3 — OpenHermes 2.5: The Shared Steering Wheel **Type**: Hub-and-spoke **Purpose**: One dataset (~1M examples, Teknium) propagated through a family of models. The course thesis ("data matters more than algorithm") made concrete. **Reading the diagram**: Center = the dataset. Spokes = the model families it steered. A flaw in the dataset would propagate to every spoke. ```mermaid flowchart TD OH["OpenHermes 2.5\nTeknium · ~1,001,551 examples\nopen-source + custom synthetic"] OH --> M1["OpenHermes 2.5 (model)"] OH --> M2["Nous Hermes 2 family"] OH --> M3["Hermes 3\n(Llama 3.1 8B/70B/405B)"] Note["THE THESIS, CONCRETE:\none steering wheel -> a family of steered models.\nFlaws propagate too. Audit your data like code."] OH -.->|"data matters more than algorithm"| Note style OH fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style M1 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style M2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style M3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Note fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4 ``` --- ## Diagram 4 — The Uncensored-in-Harness Synthesis **Type**: Comparison with annotation **Purpose**: Restate the FT00/FT23 synthesis with the concrete Dolphin/Hermes examples. The model is steered to execute; the harness bounds what it may execute. **Reading the diagram**: Bottom = the steered model (Layer 3). Top = the harness (Layer 5). The annotation is the rule: uncensor to execute, harness to bound — both required, and the harness requirement is RAISED, not lowered, by uncensoring. ```mermaid block-beta columns 1 subgraph Harness["LAYER 5 — THE HARNESS — required, raised"] H["policy gates · audit logs · threat model\nbounds what the model MAY do"] end Note["SYNTHESIS (FT00 / FT23):\nUncensor the model so it EXECUTES\n(Dolphin compliance, Hermes steerability)\nHarness the model so it executes only what it SHOULD\n(uncensoring RAISES the harness requirement)"] subgraph Model["LAYER 3 — THE STEERED MODEL"] D["Dolphin3.0-R1: compliance + R1 reasoning"] He["Hermes 3: neutral, highly steerable, unlocked"] end Model --> Note Note --> Harness style H fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Note fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4 style D fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style He fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 ``` > **Diagram 4 validation note:** The `subgraph` labels avoid embedded parentheses (block-beta prefers labels without them); an em-dash is used in the Layer 5 label instead. --- ## Validation notes - All four diagrams use the course design system colors: `#14141f` panel fill, `#5eead4` accent for primary, `rgba(94,234,212,0.3–0.5)` / `rgba(255,255,255,0.08)` for secondary borders, `#e4e4e8` / `#9494a0` for text. - Paste each into [Mermaid Live Editor](https://mermaid.live) to render. All use stable Mermaid syntax (`block-beta`, `flowchart`) supported in current Mermaid (v10.4+). - For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js.