MangaAssist Fine-Tuning Topic Scenario Map
This map connects every non-intent fine-tuning topic to concrete MangaAssist scenarios. The existing Intent-Classification folder already covers the 10-intent DistilBERT router in depth, so the companion documents below focus on the rest of the fine-tuning stack.
Shared MangaAssist Baseline
| Item | Baseline assumption |
|---|---|
| Product | MangaAssist shopping and support chatbot |
| Main router | DistilBERT intent classifier |
| Intent count | 10 known intents |
| Intent dataset | 50K production examples + 5K synthetic examples |
| Router quality | about 92.1% top-1 accuracy after fine-tuning |
| Router latency | under 15 ms P95 |
| Retrieval store | manga catalog, metadata, reviews, editorial tags, and help content |
| High-risk flows | escalation, returns, checkout, order tracking, age-sensitive recommendations |
Scenario Documents Created
| Folder | New companion document | Scenario focus |
|---|---|---|
Embedding-and-Retrieval |
02-embedding_retrieval_scenarios_mangaassist.md |
contrastive embedding adapter for manga search and recommendations |
Embedding-and-Retrieval |
03-cross_encoder_reranker_scenarios_mangaassist.md |
reranking retrieved manga results with click and editorial judgments |
Embedding-and-Retrieval |
14-raft_scenarios_mangaassist.md |
retrieval-augmented fine-tuning for grounded answers |
Fine-Tuning-Techniques |
04-lora_qlora_scenarios_mangaassist.md |
LoRA and QLoRA for self-hosted manga response models |
Fine-Tuning-Techniques |
11-prompt_prefix_tuning_scenarios_mangaassist.md |
soft prompt and prefix tuning for workflow style control |
Fine-Tuning-Techniques |
12-quantization_aware_training_scenarios_mangaassist.md |
QAT for low-latency classifiers and rerankers |
Model-Compression-Optimization |
05-knowledge_distillation_scenarios_mangaassist.md |
teacher-student compression for response and routing models |
Model-Compression-Optimization |
15-mixture_of_experts_scenarios_mangaassist.md |
expert routing by genre, task, and support workflow |
Learning-Strategies |
06-continual_learning_scenarios_mangaassist.md |
monthly updates without forgetting older manga patterns |
Learning-Strategies |
07-few_shot_learning_scenarios_mangaassist.md |
adding new intents and seasonal patterns with few labels |
Learning-Strategies |
13-multi_task_learning_scenarios_mangaassist.md |
shared model for intent, sentiment, entities, and risk |
Alignment-RLHF |
08-sentiment_classifier_scenarios_mangaassist.md |
frustration and escalation sentiment detection |
Alignment-RLHF |
10-rlhf_dpo_alignment_scenarios_mangaassist.md |
preference tuning for accurate, on-brand manga answers |
Training-Infrastructure |
09-training_mlops_scenarios_mangaassist.md |
repeatable training pipeline and release gates |
Training-Infrastructure |
16-data_curation_synthetic_generation_scenarios_mangaassist.md |
production labels, synthetic examples, and label-noise control |
Training-Infrastructure |
17-interpretability_scenarios_mangaassist.md |
inspection after fine-tuning and before promotion |
Training-Infrastructure |
18-capstone_decision_scenarios_mangaassist.md |
end-to-end decision playbook across all techniques |
Fine-Tuning-Dry-Run |
00-fine_tuning_scenario_execution_map_mangaassist.md |
execution checklist that sequences the scenario documents |
How To Use This Set
- Start with the existing intent-classification scenario docs.
- Use the embedding and reranker scenarios when MangaAssist search or recommendations feel irrelevant.
- Use LoRA, prompt tuning, DPO, RAFT, and distillation when answer quality is the bottleneck.
- Use continual learning, few-shot learning, data curation, and MLOps when the issue is update cadence.
- Use interpretability and the capstone decision guide before promoting any new model.
Decision Principle
MangaAssist should not fine-tune because a method is fashionable. Fine-tune only when the scenario has:
- measurable user or business pain,
- labeled or preference data that matches that pain,
- an offline metric tied to production behavior,
- a latency and cost budget,
- a rollback path.