Home

Transformer Training — Project Wiki

Project: Transformer Fine-Tuning and LLM Preference Alignment
Repository content: PyTorch / Hugging Face DistilBERT classification plus a separate, unfinished Phi-2 SFT/DPO example. It will be introduced in the next iteration.

Current status: DistilBERT script and notebook are syntactically valid; end-to-end training was not tested for this documentation package. PPO is not implemented. ORPO example. It will be implemented in the next iteration.

Start here

• Quick-Start — install and run the available classifier
• System-Architecture — implemented HLD vs proposed preference-training workflow
• Low-Level-Design — functions, dependencies, and control flow
• DistilBERT-Workflow and Model-Internals — actual classification pipeline
• Phi2-SFT-DPO-Design — non-executable design-stage alignment example
• Known-Issues and Validation-Report — limitations and tested claims

At a glance

flowchart LR
    Dev[Developer / Researcher] --> CLI[Python CLI]
    Dev --> NB[Jupyter / Colab Notebook]
    CLI --> Pipeline[DistilBERT Training Workflow]
    NB --> Pipeline
    HF[(Hugging Face Model Hub)] -->|download tokenizer and pretrained weights| Pipeline
    Pipeline --> Models[(Local Model Checkpoints)]
    Pipeline --> Metrics[(Validation Reports / Logs)]
    Dev -.-> Draft[Phi-2 SFT / DPO Draft: not runnable]
    Draft -.-> HF

Loading

The application runs locally or in a notebook. This repository includes no deployed web API, microservice, RAG retrieval system, or autonomous agent.

Source files

File	Purpose	Status
DistilBERT_Colab_Fixed_2026.py	Four-class support-ticket classification, training, inference and frozen-backbone comparison	Parses successfully
DistilBERT_Colab_Fixed_2026.ipynb	Jupyter equivalent with executable cells	Notebook parses and code cells compile
DPO_SFT.py	Phi-2 + LoRA SFT / preference optimization draft
requirements.txt	Combined optional dependency list	Installation not tested
README.md	Repository guidance	Updated; original preserved under docs/ORIGINAL_README.md

All source code remains in the main GitHub repository; this Wiki lives in a separate .wiki.git repository.