Skip to content

Training and Checkpoints

Sudhanshu edited this page Nov 11, 2026 · 1 revision

Training and Artifacts

stateDiagram-v2
    [*] --> Validate
    Validate --> Training : dependencies and model available
    Validate --> [*] : validate-data mode
    Training --> QwenAdapter : Qwen LoRA
    Training --> SFTAdapter : legacy SFT
    SFTAdapter --> DPOAdapter : reference frozen
    Training --> ORPOAdapter : separate legacy run
    QwenAdapter --> [*]
    DPOAdapter --> [*]
    ORPOAdapter --> [*]
Loading

Qwen

Trainer.train() runs model loss/backpropagation. The code uses save_strategy="no" and explicitly saves the final adapter and tokenizer to qwen_alien_romulus_finetuned/. It also saves training metrics, but no held-out validation or intermediate checkpoint is configured.

Legacy

run_sft_dpo() saves sft_adapter/ and dpo_adapter/. run_orpo() saves orpo_adapter/. For inference the LoRA adapter must be loaded against the matching base model. Legacy generator outputs in the CLI are qualitative comparisons, not an evaluation report.

Do not commit trained weights unintentionally

The repo .gitignore excludes output folders and common checkpoint patterns. Inspect git status before pushing; never add API keys, tokens or private data.

Clone this wiki locally