# Diagrams — Deep-Dive FTDD-01: MiniCPM Family (OpenBMB) **Deep-Dive**: FTDD-01 — MiniCPM Family (OpenBMB) **Diagram count**: 4 **Tool**: Mermaid (primary). Each diagram validated in [Mermaid Live Editor](https://mermaid.live). --- ## Diagram 1 — The MiniCPM Family (modality axis) **Type**: Linear family tree **Purpose**: The single diagram that maps the four MiniCPM variants to the modality gap each fills. Read it as a progression: text → denser text → +vision → +audio/full-duplex. **Reading the diagram**: Left = the 1B on-ramp. Each step right adds either capability or modality. SigLip-400M is the load-bearing component that turns a text base into MiniCPM-V. ```mermaid flowchart LR A["MiniCPM5-1B\n~1.08B · text\non-ramp hero"] B["MiniCPM3-4B\n4B · text\nmid-size reasoning"] C["MiniCPM-V 4.6\n+ SigLip-400M\nvision-language"] D["MiniCPM-o 4.5\nomni-modal\nfull-duplex streaming"] A -->|"adds: density"| B B -->|"adds: vision encoder"| C C -->|"adds: audio + duplex"| D style A fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style B fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style C fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style D fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 ``` --- ## Diagram 2 — The Ultra\* Datasets → Steering Layer Map **Type**: Mapping (dataset → stack layer → course module) **Purpose**: Show why the Ultra\* datasets are the course's data reference. Each dataset maps to a specific steering layer (FT00's stack) and a specific course module — open data lets you point at the signal. **Reading the diagram**: Three rows, one per dataset. Each row names the dataset, the steering technique it enables, and the course module that uses it as the worked example. ```mermaid flowchart TB subgraph DATA["THE Ultra* OPEN DATASETS"] UC["UltraChat\nmulti-turn dialogue"] UF["UltraFeedback\npreference / feedback pairs\n(arXiv:2310.01377)"] UFW["Ultra-FineWeb\ncurated web pretraining mix"] end subgraph STEER["STEERING LAYER (FT00 stack)"] S1["Layer 3 — SFT\ninstruction format"] S2["Layer 3 — DPO\npreference signal"] S3["Layer 1 / CPT\nbase distribution"] end subgraph MOD["COURSE MODULE"] M1["FT04, FT12"] M2["FT05, FT13"] M3["FT06"] end UC --> S1 --> M1 UF --> S2 --> M2 UFW --> S3 --> M3 style DATA fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style STEER fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style MOD fill:#08080c,stroke:rgba(255,255,255,0.12),color:#9494a0 style UC fill:#08080c,stroke:rgba(94,234,212,0.4),color:#e4e4e8 style UF fill:#08080c,stroke:rgba(94,234,212,0.4),color:#e4e4e8 style UFW fill:#08080c,stroke:rgba(94,234,212,0.4),color:#e4e4e8 ``` --- ## Diagram 3 — Why MiniCPM Is the Teaching Vehicle (three properties) **Type**: Three-pillar **Purpose**: The three properties that make MiniCPM the course's default base, each mapped to the course need it satisfies. This is the diagram that justifies the choice of base pedagogically. **Reading the diagram**: Three pillars supporting one claim — "the ideal teaching base." Each pillar is a property (cheap iteration, auditable, Apache-2.0) and each maps to a concrete course requirement. ```mermaid flowchart TB P1["CHEAP ITERATION\n1B fine-tunes in minutes\non a consumer GPU"] P2["AUDITABLE PROVENANCE\nopen weights + data + recipe\npoint at every byte"] P3["APACHE-2.0\nno MAU / field-of-use\nno license friction"] CLAIM["THE IDEAL TEACHING BASE\na model you can break,\nreset, and break again"] P1 --> CLAIM P2 --> CLAIM P3 --> CLAIM style P1 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style P2 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style P3 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style CLAIM fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 ``` --- ## Diagram 4 — SWIFT vs LLaMA-Factory (the fine-tuning decision) **Type**: Two-track comparison **Purpose**: The decision diagram for choosing a MiniCPM fine-tuning framework. The deciding factor is modality — text models are a toss-up; vision/omni models favor SWIFT. **Reading the diagram**: Top track = SWIFT (first-party). Bottom track = LLaMA-Factory (general). The annotation marks the deciding factor. ```mermaid flowchart TB subgraph SWIFT["SWIFT (ModelBest / OpenBMB — first-party)"] direction TB S1["Native support for ALL MiniCPM variants"] S2["First-party chat templates & modality handlers"] S3["Center of gravity = OpenBMB ecosystem"] S1 --> S2 --> S3 end subgraph LF["LLaMA-Factory (general-purpose)"] direction TB L1["Unified interface for dozens of model families"] L2["Add MiniCPM as a config change in a multi-model pipeline"] L3["Vision/omni support lags SWIFT by weeks-months"] L1 --> L2 --> L3 end DECISION{"Base modality?\ntext vs vision/omni"} DECISION -->|"MiniCPM-V / MiniCPM-o\n(first-party modality)"| SWIFT DECISION -->|"MiniCPM5-1B / 3-4B\n(multi-model pipeline)"| LF style SWIFT fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style LF fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style DECISION fill:#08080c,stroke:rgba(240,168,104,0.5),stroke-dasharray: 4 2,color:#f0a868 style S1 fill:#08080c,stroke:rgba(94,234,212,0.3),color:#e4e4e8 style S2 fill:#08080c,stroke:rgba(94,234,212,0.3),color:#e4e4e8 style S3 fill:#08080c,stroke:rgba(94,234,212,0.3),color:#e4e4e8 style L1 fill:#08080c,stroke:rgba(94,234,212,0.3),color:#e4e4e8 style L2 fill:#08080c,stroke:rgba(94,234,212,0.3),color:#e4e4e8 style L3 fill:#08080c,stroke:rgba(240,168,104,0.3),color:#f0a868 ``` --- ## Validation notes - All four diagrams use the course design system colors: `#14141f` panel fill, `#5eead4` accent for primary, `rgba(255,255,255,0.12)` for secondary borders, `#e4e4e8` / `#9494a0` for text. The warn tone (`#f0a868`) marks the decision node and the lagging-support caveat. - Paste each into [Mermaid Live Editor](https://mermaid.live) to render. All use stable Mermaid syntax (`flowchart LR/TB`, `subgraph`) 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.