src/make_mamba_finetuning_pptx.py

import os
from pptx import Presentation
from pptx.util import Inches, Pt
from pptx.enum.text import PP_ALIGN
from pptx.dml.color import RGBColor
from pptx.enum.shapes import MSO_SHAPE

def create_proof_of_work_presentation():
    pptx_path = "/home/adminuser/aiops_pocs/MAMBA_FINETUNING/docs/MAMBA_FINETUNING_PROOF_OF_WORK_PRESENTATION.pptx"

    prs = Presentation()
    prs.slide_width = Inches(13.333)
    prs.slide_height = Inches(7.5)

    # Ultra-Premium Dark Theme Color Palette
    COLOR_BG = RGBColor(11, 15, 25)         # Deep Obsidian Slate
    COLOR_CARD = RGBColor(30, 41, 59)       # Slate 800
    COLOR_CARD_DARK = RGBColor(20, 27, 44)  # Slate 900
    COLOR_CYAN = RGBColor(6, 182, 212)      # Electric Cyan
    COLOR_GREEN = RGBColor(16, 185, 129)    # Emerald Green
    COLOR_PURPLE = RGBColor(168, 85, 247)   # Vivid Purple
    COLOR_AMBER = RGBColor(245, 158, 11)    # Radiant Amber
    COLOR_TEXT = RGBColor(248, 250, 252)    # Pure White Slate
    COLOR_MUTED = RGBColor(148, 163, 184)   # Slate Muted

    def apply_bg(slide):
        bg = slide.shapes.add_shape(MSO_SHAPE.RECTANGLE, 0, 0, Inches(13.333), Inches(7.5))
        bg.fill.solid()
        bg.fill.fore_color.rgb = COLOR_BG
        bg.line.fill.background()

    def add_header(slide, title_text, chapter_text="MY PROOF OF WORK"):
        tb = slide.shapes.add_textbox(Inches(0.8), Inches(0.4), Inches(11.7), Inches(0.9))
        tf = tb.text_frame
        tf.word_wrap = True
        
        p_cat = tf.paragraphs[0]
        p_cat.text = chapter_text.upper()
        p_cat.font.size = Pt(11)
        p_cat.font.bold = True
        p_cat.font.color.rgb = COLOR_CYAN
        
        p_title = tf.add_paragraph()
        p_title.text = title_text
        p_title.font.size = Pt(24)
        p_title.font.bold = True
        p_title.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 1: Title Slide (First Person Proof of Work)
    # -------------------------------------------------------------
    slide1 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide1)
    
    # Hero Card Background Container
    hero_card = slide1.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(0.8), Inches(1.2), Inches(11.7), Inches(5.1))
    hero_card.fill.solid()
    hero_card.fill.fore_color.rgb = COLOR_CARD_DARK
    hero_card.line.color.rgb = COLOR_GREEN
    hero_card.line.width = Pt(2)
    
    # Top Accent Line
    line = slide1.shapes.add_shape(MSO_SHAPE.RECTANGLE, Inches(0.8), Inches(1.2), Inches(11.7), Inches(0.12))
    line.fill.solid()
    line.fill.fore_color.rgb = COLOR_GREEN
    line.line.fill.background()
    
    tb = slide1.shapes.add_textbox(Inches(1.2), Inches(1.6), Inches(10.9), Inches(4.2))
    tf = tb.text_frame
    tf.word_wrap = True
    
    p1 = tf.paragraphs[0]
    p1.text = "🏆 PROOF OF WORK: LLM FINE-TUNING & EXPERIMENTATION"
    p1.font.size = Pt(13)
    p1.font.bold = True
    p1.font.color.rgb = COLOR_GREEN
    
    p2 = tf.add_paragraph()
    p2.text = "How I Fine-Tuned Mamba 7B for Deep CoT Reasoning"
    p2.font.size = Pt(34)
    p2.font.bold = True
    p2.font.color.rgb = COLOR_TEXT
    
    p3 = tf.add_paragraph()
    p3.text = "\nA first-person technical walkthrough detailing how I transformed a Selective State Space Model into a high-reasoning engine by instilling DeepSeek-R1 style Chain-of-Thought (<think>) capabilities on 2x NVIDIA H200 GPUs."
    p3.font.size = Pt(15)
    p3.font.color.rgb = COLOR_MUTED

    # Stat Badges at Bottom
    badges = [
        ("AUTHOR", "Naman Adep", COLOR_GREEN),
        ("DATASET", "16,710 CoT Samples", COLOR_CYAN),
        ("GPU CLUSTER", "2x NVIDIA H200 NVL", COLOR_PURPLE),
        ("COT TRIGGER", "100% Success Rate", COLOR_AMBER)
    ]
    for idx, (b_title, b_val, b_color) in enumerate(badges):
        x = Inches(1.2 + idx * 2.7)
        b_box = slide1.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, x, Inches(4.8), Inches(2.5), Inches(1.1))
        b_box.fill.solid()
        b_box.fill.fore_color.rgb = COLOR_CARD
        b_box.line.color.rgb = b_color
        b_box.line.width = Pt(1)
        
        b_tf = b_box.text_frame
        b_tf.word_wrap = True
        p_t = b_tf.paragraphs[0]
        p_t.text = b_title
        p_t.font.size = Pt(9)
        p_t.font.bold = True
        p_t.font.color.rgb = COLOR_MUTED
        
        p_v = b_tf.add_paragraph()
        p_v.text = b_val
        p_v.font.size = Pt(12)
        p_v.font.bold = True
        p_v.font.color.rgb = b_color

    # -------------------------------------------------------------
    # SLIDE 2: Executive Summary & Proof of Work
    # -------------------------------------------------------------
    slide2 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide2)
    add_header(slide2, "Executive Summary: What I Accomplished", "My Proof of Work")
    
    pillars = [
        ("1. Data Pipeline Creation", "I processed 16,710 DeepSeek-R1 distilled reasoning samples into structured multi-turn conversation formats with explicit <think>...</think> tags.", COLOR_CYAN),
        ("2. Multi-GPU Fine-Tuning", "I adapted Mamba 7B's linear projection layers (in_proj, x_proj, dt_proj) using LoRA (rank=16) on 2x NVIDIA H200 NVL GPUs.", COLOR_GREEN),
        ("3. 50-Prompt Evaluation", "I built an automated evaluation harness testing 50 complex technical prompts across Math, Systems, Security, and AI Theory.", COLOR_PURPLE)
    ]
    
    for idx, (title, desc, color) in enumerate(pillars):
        x = Inches(0.8 + idx * 3.9)
        card = slide2.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, x, Inches(1.6), Inches(3.6), Inches(5.2))
        card.fill.solid()
        card.fill.fore_color.rgb = COLOR_CARD
        card.line.color.rgb = color
        card.line.width = Pt(2)
        
        bar = slide2.shapes.add_shape(MSO_SHAPE.RECTANGLE, x, Inches(1.6), Inches(3.6), Inches(0.1))
        bar.fill.solid()
        bar.fill.fore_color.rgb = color
        bar.line.fill.background()
        
        tf = card.text_frame
        tf.word_wrap = True
        
        p = tf.paragraphs[0]
        p.text = title
        p.font.size = Pt(17)
        p.font.bold = True
        p.font.color.rgb = color
        
        p_d = tf.add_paragraph()
        p_d.text = f"\n{desc}"
        p_d.font.size = Pt(13)
        p_d.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 3: Why Fine-Tuning Mamba Matters
    # -------------------------------------------------------------
    slide3 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide3)
    add_header(slide3, "Why Fine-Tuning Mamba is a Game-Changer", "Architectural Vision")
    
    c1 = slide3.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(0.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c1.fill.solid()
    c1.fill.fore_color.rgb = COLOR_CARD
    c1.line.color.rgb = COLOR_AMBER
    c1.line.width = Pt(2)
    tf1 = c1.text_frame
    tf1.word_wrap = True
    
    p = tf1.paragraphs[0]
    p.text = "⚠️ The Base Mamba Limitation"
    p.font.size = Pt(18)
    p.font.bold = True
    p.font.color.rgb = COLOR_AMBER
    
    b1 = [
        "• Base Mamba is blazingly fast but omits intermediate reasoning steps.",
        "• It jumps directly to final solutions, occasionally missing nuanced edge cases or intermediate algebraic derivations.",
        "• Lacks native Chain-of-Thought (<think>) internal reasoning traces."
    ]
    for b in b1:
        p = tf1.add_paragraph()
        p.text = b
        p.font.size = Pt(13)
        p.font.color.rgb = COLOR_TEXT

    c2 = slide3.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(6.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c2.fill.solid()
    c2.fill.fore_color.rgb = COLOR_CARD
    c2.line.color.rgb = COLOR_GREEN
    c2.line.width = Pt(2)
    tf2 = c2.text_frame
    tf2.word_wrap = True
    
    p = tf2.paragraphs[0]
    p.text = "✨ What My Fine-Tuning Unlocks"
    p.font.size = Pt(18)
    p.font.bold = True
    p.font.color.rgb = COLOR_GREEN
    
    b2 = [
        "• Instills systematic step-by-step deconstruction before generating answers.",
        "• Combines Transformer-level reasoning quality with Mamba's constant O(1) memory efficiency.",
        "• Enables zero KV-cache VRAM inflation during 256k long-context reasoning."
    ]
    for b in b2:
        p = tf2.add_paragraph()
        p.text = b
        p.font.size = Pt(13)
        p.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 4: Quantitative Transformation Metrics
    # -------------------------------------------------------------
    slide4 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide4)
    add_header(slide4, "Empirical Results: Before vs After Fine-Tuning", "Quantitative Metrics")
    
    stats = [
        ("100%", "COT TRIGGER RATE", "Every query post-fine-tuning initiates explicit <think> reasoning steps.", COLOR_GREEN),
        ("1.85x", "CONTENT DENSITY", "Output expanded from 972 chars to 1,794 chars average due to detailed step derivations.", COLOR_CYAN),
        ("3.52s", "AVERAGE LATENCY", "Maintains sub-4-second response completion time while adding deep reasoning.", COLOR_PURPLE),
        ("O(1)", "VRAM EFFICIENCY", "Zero memory inflation during processing, retaining fixed recurrent state buffer.", COLOR_AMBER)
    ]
    
    for idx, (num, label, desc, color) in enumerate(stats):
        row = idx // 2
        col = idx % 2
        x = Inches(0.8 + col * 5.9)
        y = Inches(1.6 + row * 2.6)
        
        card = slide4.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, x, y, Inches(5.6), Inches(2.3))
        card.fill.solid()
        card.fill.fore_color.rgb = COLOR_CARD
        card.line.color.rgb = color
        card.line.width = Pt(1.5)
        
        tf = card.text_frame
        tf.word_wrap = True
        
        p_num = tf.paragraphs[0]
        p_num.text = num
        p_num.font.size = Pt(36)
        p_num.font.bold = True
        p_num.font.color.rgb = color
        
        p_lbl = tf.add_paragraph()
        p_lbl.text = label
        p_lbl.font.size = Pt(11)
        p_lbl.font.bold = True
        p_lbl.font.color.rgb = COLOR_MUTED
        
        p_desc = tf.add_paragraph()
        p_desc.text = desc
        p_desc.font.size = Pt(12)
        p_desc.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 5: Transformation Breakdown (Before vs After)
    # -------------------------------------------------------------
    slide5 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide5)
    add_header(slide5, "Qualitative Comparison: Response Evolution", "Side-by-Side Case Study")
    
    c1 = slide5.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(0.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c1.fill.solid()
    c1.fill.fore_color.rgb = COLOR_CARD
    c1.line.color.rgb = RGBColor(239, 68, 68)
    tf1 = c1.text_frame
    tf1.word_wrap = True
    
    p = tf1.paragraphs[0]
    p.text = "🔴 Base Mamba 7B (Before Fine-Tuning)"
    p.font.size = Pt(17)
    p.font.bold = True
    p.font.color.rgb = RGBColor(239, 68, 68)
    
    b1 = [
        "Prompt: Solve 3x² + 14x - 5 = 0\n",
        "Output:",
        "3x² + 14x - 5 = 0",
        "(3x - 1)(x + 5) = 0",
        "x = 1/3 or x = -5",
        "\n• Characteristic: Direct answer, skipped factorization steps, no verification."
    ]
    for b in b1:
        p = tf1.add_paragraph()
        p.text = b
        p.font.size = Pt(12)
        p.font.color.rgb = COLOR_TEXT

    c2 = slide5.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(6.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c2.fill.solid()
    c2.fill.fore_color.rgb = COLOR_CARD
    c2.line.color.rgb = COLOR_GREEN
    tf2 = c2.text_frame
    tf2.word_wrap = True
    
    p = tf2.paragraphs[0]
    p.text = "🟢 Fine-Tuned Mamba 7B (After Fine-Tuning)"
    p.font.size = Pt(17)
    p.font.bold = True
    p.font.color.rgb = COLOR_GREEN
    
    b2 = [
        "Output:",
        "<think>",
        "Step 1: Identify coefficients a=3, b=14, c=-5.",
        "Step 2: Find numbers multiplying to -15 & adding to 14 (15 and -1).",
        "Step 3: Factor by grouping: 3x(x+5) - 1(x+5) = (3x-1)(x+5).",
        "Step 4: Verify via quadratic formula: x = (-14 ± 16) / 6.",
        "</think>",
        "\nFinal Answer: x = 1/3, x = -5."
    ]
    for b in b2:
        p = tf2.add_paragraph()
        p.text = b
        p.font.size = Pt(12)
        p.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 6: Technical Implementation Highlights
    # -------------------------------------------------------------
    slide6 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide6)
    add_header(slide6, "Technical Implementation & LoRA Architecture", "Fine-Tuning Setup")
    
    box = slide6.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(0.8), Inches(1.6), Inches(11.7), Inches(5.2))
    box.fill.solid()
    box.fill.fore_color.rgb = COLOR_CARD
    box.line.color.rgb = COLOR_CYAN
    tf = box.text_frame
    tf.word_wrap = True
    
    p = tf.paragraphs[0]
    p.text = "LoRA Target Modules & Hyperparameter Configuration"
    p.font.size = Pt(16)
    p.font.bold = True
    p.font.color.rgb = COLOR_CYAN
    
    code_text = """# My PyTorch & PEFT LoRA Configuration for Falcon Mamba 7B
peft_config = LoraConfig(
    r=16,                                    # LoRA Rank
    lora_alpha=32,                           # Alpha Scaling Factor
    target_modules=["in_proj", "x_proj", "dt_proj"], # SSM Linear Projection Layers
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.CAUSAL_LM
)

# Training Parameters on 2x NVIDIA H200 NVL GPUs
training_args = TrainingArguments(
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,                      # Cosine Learning Rate Schedule
    bf16=True,                               # Mixed Precision bfloat16
    num_train_epochs=3
)"""
    
    p_code = tf.add_paragraph()
    p_code.text = code_text
    p_code.font.size = Pt(12)
    p_code.font.name = "Courier New"
    p_code.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 7: Deliverables & Proof of Work Artifacts
    # -------------------------------------------------------------
    slide7 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide7)
    add_header(slide7, "My Deliverables & Proof of Work Artifacts", "Code & Data")
    
    artifacts = [
        ("📁 Workspace Directory", "/home/adminuser/MAMBA_FINETUNING/", "Clean modular layout with docs/, src/, and data/ subdirectories.", COLOR_CYAN),
        ("📊 50-Prompt Report", "docs/50_PROMPTS_MAMBA_BASE_VS_REASONING_COMPARISON.md", "215 KB side-by-side empirical benchmark report.", COLOR_GREEN),
        ("🐍 Training Pipeline", "src/train_mamba_reasoning.py & prepare_reasoning_dataset.py", "End-to-end dataset formatter & multi-GPU trainer.", COLOR_PURPLE),
        ("🎨 Presentation Decks", "docs/*.pptx", "Visual PowerPoint presentations ready for technical review.", COLOR_AMBER)
    ]
    
    for idx, (title, path, desc, color) in enumerate(artifacts):
        row = idx // 2
        col = idx % 2
        x = Inches(0.8 + col * 5.9)
        y = Inches(1.6 + row * 2.6)
        
        card = slide7.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, x, y, Inches(5.6), Inches(2.3))
        card.fill.solid()
        card.fill.fore_color.rgb = COLOR_CARD
        card.line.color.rgb = color
        card.line.width = Pt(1.5)
        
        tf = card.text_frame
        tf.word_wrap = True
        
        p_t = tf.paragraphs[0]
        p_t.text = title
        p_t.font.size = Pt(16)
        p_t.font.bold = True
        p_t.font.color.rgb = color
        
        p_p = tf.add_paragraph()
        p_p.text = path
        p_p.font.size = Pt(11)
        p_p.font.bold = True
        p_p.font.name = "Courier New"
        p_p.font.color.rgb = COLOR_MUTED
        
        p_d = tf.add_paragraph()
        p_d.text = desc
        p_d.font.size = Pt(12)
        p_d.font.color.rgb = COLOR_TEXT

    # -------------------------------------------------------------
    # SLIDE 8: Conclusion & Summary
    # -------------------------------------------------------------
    slide8 = prs.slides.add_slide(prs.slide_layouts[6])
    apply_bg(slide8)
    add_header(slide8, "Conclusion: The Power of Fine-Tuned SSMs", "Final Summary")
    
    c1 = slide8.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(0.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c1.fill.solid()
    c1.fill.fore_color.rgb = COLOR_CARD
    c1.line.color.rgb = COLOR_GREEN
    tf1 = c1.text_frame
    tf1.word_wrap = True
    
    p = tf1.paragraphs[0]
    p.text = "🎯 Core Breakthrough"
    p.font.size = Pt(18)
    p.font.bold = True
    p.font.color.rgb = COLOR_GREEN
    
    b1 = [
        "• Fine-tuning proves that Mamba models can master deep, multi-step Chain-of-Thought (<think>) reasoning.",
        "• Bridges the reasoning gap between Transformers and State Space Models.",
        "• Delivers high-level mathematical and algorithmic correctness without sacrificing sub-4-second response latency."
    ]
    for b in b1:
        p = tf1.add_paragraph()
        p.text = b
        p.font.size = Pt(13)
        p.font.color.rgb = COLOR_TEXT

    c2 = slide8.shapes.add_shape(MSO_SHAPE.ROUNDED_RECTANGLE, Inches(6.8), Inches(1.6), Inches(5.6), Inches(5.2))
    c2.fill.solid()
    c2.fill.fore_color.rgb = COLOR_CARD
    c2.line.color.rgb = COLOR_CYAN
    tf2 = c2.text_frame
    tf2.word_wrap = True
    
    p = tf2.paragraphs[0]
    p.text = "🚀 Future Applications"
    p.font.size = Pt(18)
    p.font.bold = True
    p.font.color.rgb = COLOR_CYAN
    
    b2 = [
        "1. Real-Time IDE Autocomplete with CoT context validation.",
        "2. Long-Context Document Reasoning (256k tokens) with zero KV-cache memory crashes.",
        "3. High-throughput edge server deployments running lightweight reasoning models."
    ]
    for b in b2:
        p = tf2.add_paragraph()
        p.text = b
        p.font.size = Pt(13)
        p.font.color.rgb = COLOR_TEXT

    prs.save(pptx_path)
    print("Successfully created First-Person Proof of Work Mamba presentation:", pptx_path)

if __name__ == "__main__":
    create_proof_of_work_presentation()