Fine-Tuning Foundational Models for MangaAssist

Overview

This folder contains 18 deep-dive scenario documents covering fine-tuning, model customization, post-fine-tuning inspection techniques, and capstone intuition synthesis relevant to the MangaAssist chatbot. Each document is written as a group discussion among 5 engineers, includes mathematical derivations with geometric intuition, layer-by-layer mermaid diagrams or inspection workflows showing what happens inside models during training, production-grade code, and references to foundational research papers.

MangaAssist Scenario Companion Set

For the newer MangaAssist-grounded scenario documents, start with:

• MangaAssist Fine-Tuning Topic Scenario Map
• Fine-Tuning Scenario Execution Map

These companion docs sit beside the original deep dives and cover the non-intent fine-tuning topics: embeddings, reranking, RAFT, LoRA/QLoRA, prompt tuning, QAT, distillation, MoE, continual learning, few-shot learning, multi-task learning, sentiment, DPO/RLHF, MLOps, data curation, interpretability, and capstone decision-making.

Reading Order

Tier 1: Start Here (Core Fine-Tuning for MangaAssist Production)

These are the techniques already deployed or planned for MangaAssist:

Tier 2: Advanced Model Customization

These extend MangaAssist with more sophisticated fine-tuning:

Tier 3: Specialized Techniques

For specific production challenges:

Tier 4: Infrastructure and Data

Tier 5: Capstone Synthesis

Group Discussion Personas

Every document features debates among these five engineers, representing a real Amazon-style cross-functional team:

Discussion Format

### Decision Point: Should we use LoRA rank 8 or 16?

**Priya (ML Engineer):** Rank 16 gives us more expressive capacity to capture
manga-specific patterns. The SVD analysis of weight updates shows the top 16
singular values capture 94% of the variance, vs 87% for rank 8...

**Marcus (Architect):** But rank 16 doubles adapter memory from 4MB to 8MB per
adapter. With 10 concurrent adapters for different intents, that's 80MB just
for adapters...

**Aiko (Data Scientist):** The ablation study on our validation set shows:
rank 4 = 89.1%, rank 8 = 91.8%, rank 12 = 92.3%, rank 16 = 92.5%.
Diminishing returns above rank 8...

**Sam (PM):** The cost difference is $240/month (rank 8) vs $480/month (rank 16)
for GPU memory. For 0.7% accuracy gain, that's $342 per quality point...

**Jordan (MLOps):** More parameters means longer CI validation cycles. Rank 8
validates in 12 minutes, rank 16 takes 22 minutes...

> **Resolution:** Rank 8 chosen. The 91.8% accuracy meets our 90% threshold,
> the cost-per-quality-point for rank 16 ($342/point) exceeds our $300
> threshold, and faster CI cycles matter for weekly adapter updates.

Mathematical Prerequisites

To get the most from these documents, you should be comfortable with:

Master Research Paper Reading List

Foundational (Read These First)

Fine-Tuning Techniques

Contrastive Learning and Retrieval

Alignment and Preference Learning

Distillation and Compression

Continual and Few-Shot Learning

Multi-Task and Expert Models

RAG and Retrieval-Augmented Training

Data Quality and Synthesis

Infrastructure

Cross-References to Other Project Folders

How to Use These Documents

1. For interview prep: Start with the README, then read Tier 1 docs, then use the decision-point sections plus Docs 17 and 18 to rehearse architecture, training, interpretability, and strategic tradeoffs.
2. For implementation: Read the relevant scenario doc end-to-end, then reference Doc 09 (MLOps) for deployment.
3. For understanding the math: Each doc is self-contained. Read the "Mathematical Foundations" section, then trace through the mermaid layer diagrams.
4. For the group discussion format: Each decision point shows how a real team would debate the tradeoff. Use these as templates for your own technical discussions, then read Doc 18 for the cross-technique synthesis.