Replace static plan of action image with dynamic mermaid file

[INF800]

Current commit replaces the plan of action PNG file with below mermaid file:

flowchart LR
    subgraph one["Step 1"]
        A1[Deepseek-R1]:::modelStyle --> B1[Distilled reasoning data]:::dataStyle
        C1[Instruct Model]:::modelStyle --> D1[SFT]:::processStyle
        B1 --> D1
        D1 --> E1[Open R1-Distill]:::modelStyle
    end

    subgraph two["Step 2"]
        B2[RL reasoning data]:::dataStyle --> D2[GRPO with verifiable rewards]:::processStyle
        C2[Base Model]:::modelStyle --> D2
        D2 --> E2[Open R1-Zero]:::modelStyle
    end

    subgraph three["Step 3"]
        A3[Open R1-Zero]:::modelStyle --> B3[SFT reasoning data]:::dataStyle
        C3[Base Model]:::modelStyle --> D3[SFT]:::processStyle
        B3 --> D3
        D3 --> E3[GRPO with verifiable rewards]:::processStyle
        F3[RL reasoning data]:::dataStyle --> E3
        E3 --> G3[Open R1]:::modelStyle
    end

    one --> two --> three

    classDef modelStyle fill:#ffb6c1,stroke:#333,stroke-width:2px,color:#000000
    classDef dataStyle fill:#ffeb99,stroke:#333,stroke-width:2px,color:#000000
    classDef processStyle fill:#87ceeb,stroke:#333,stroke-width:2px,color:#000000

Loading

Need: To be able to make changes at code level and add more sophisticated diagrams like this one: (easy to see <-> easy to understand)

flowchart TD
    %% Define styles with darker font colors
    classDef blueBox fill:#e6f3ff,stroke:#000,stroke-width:1px,color:#000000
    classDef yellowBox fill:#fffbe6,stroke:#000,stroke-width:1px,color:#000000
    classDef pinkBox fill:#ffe6e6,stroke:#000,stroke-width:1px,color:#000000

    %% Base model
    A["DeepSeek-V3 Base<br/>(671B/37B Activated)"]:::blueBox

    %% First branch - Supervised Fine-Tuning
    B["Supervised<br/>Fine-Tuning<br/>(SFT)"]:::yellowBox
    C[("Cold Start<br/>Long CoT Data<br/>(~k samples)")]:::pinkBox

    %% RORL and CoT components
    D["Reasoning Oriented RL<br/>GRPO<br/>Rule-based Reward<br/>(Accuracy, Formatting)"]:::yellowBox
    E["+ CoT Language<br/>Consistency Reward"]:::yellowBox

    %% Middle section
    F["DeepSeek-V3 Base<br/>+ CS SFT + RORL"]:::blueBox
    G["Reasoning Prompts +<br/>Rejection Sampling<br/>(Rule-based &<br/>DS-V3 as judge)"]:::yellowBox
    H[("Reasoning Data<br/>(600k samples)")]:::pinkBox

    %% Right branch
    I["DeepSeek-V3<br/>(671B/37B Activated)"]:::blueBox
    J["CoT Prompting"]:::yellowBox
    K[("Non-Reasoning<br/>Data<br/>(200k samples)")]:::pinkBox

    %% Model variants
    L[("DeepSeek-V3<br/>SFT Data")]:::pinkBox
    M[("Combined<br/>SFT Data<br/>(800k samples)")]:::pinkBox

    %% Bottom section - Models and training
    N["Qwen2.5-Math-1.5B"]:::blueBox
    O["Qwen2.5-Math-7B"]:::blueBox
    P["Qwen2.5 14B"]:::blueBox
    Q["Qwen2.5 32B"]:::blueBox
    R["Llama-3.3-70B-Instruct"]:::blueBox
    S["Llama-3.1-8B"]:::blueBox

    T["SFT<br/>2 epochs<br/>800k samples"]:::yellowBox
    U["SFT<br/>2 epochs<br/>800k samples"]:::yellowBox
    V["RL<br/>Reasoning + Preference Reward<br/>Diverse Training Prompts"]:::yellowBox

    %% Final models
    W["DeepSeek-R1-Zero"]:::blueBox
    X["DeepSeek-R1-Distill-(Qwen/Llama)-{*B}"]:::blueBox
    Y["DeepSeek-R1"]:::blueBox

    %% Connections
    A --> B
    B --> C
    A --> D
    A --> E
    D & E --> F
    F --> G
    G --> H
    I --> J
    J --> K
    H --> M
    K --> M
    L --> M
    N & O & P & Q & R & S --> T
    T --> X
    M --> T
    M --> U
    U --> V
    V --> Y
    A --> W

    %% Subgraph for distillation
    subgraph Distillation
        N
        O
        P
        Q
        R
        S
        T
        X
    end

Loading