continualcode

Torch-native SDPO benchmark training with optional Modal-backed train steps.

Train

python -m continualcode model_name=Qwen/Qwen3-4B-Instruct-2507

Eval

python -m continualcode.benchmarks.lcb_eval split=test max_samples=100

Artifacts

• metrics.jsonl
• samples.jsonl
• checkpoint folders in configured checkpoint_dir

Core modules

• continualcode/benchmarks/auto_train.py
• continualcode/modal_train.py
• continualcode/model_utils.py
• continualcode/sdpo_loss.py