Workflow Comparison

Comparing SFT, DPO, ORPO and LoRA

Concept	In this repository	What it does
LoRA	Qwen + legacy	Parameter-efficient adaptation by low-rank trainable updates
SFT	Legacy, stage 1	Learns desired completions from prompt/answer pairs
DPO	Legacy, stage 2	Optimizes chosen responses relative to rejected responses and frozen SFT reference
ORPO	Legacy alternative mode	Combines supervised likelihood and odds-ratio preference component in a separate run
PPO	Not implemented	Would require a different policy optimization workflow
Pretraining from scratch	Not implemented	Would require large-scale corpus, tokenizer/model training, and suitable compute

The Qwen example is an instruction LoRA training demonstration, with causal LM loss applied to both prompt and answer tokens. No RAG, MCP, agent framework, online serving API, or distributed data parallelism is implemented.