Check readiness
Confirm the dataset is clean, pinned and free of evaluation overlap before any GPU starts.
ML · ships today
Fine-tune and
evaluate an LLM
Train on your GPUs, score every checkpoint, promote with approval.
Ships as training/e2e-finetune
Check the data, fine-tune through the engine your environment provides, score each checkpoint on held-out data, then wait for a person to approve the release. Preference alignment and serving continue from here.
The pipeline
Train
Fine-tune on your GPUs through Metaflow, SkyPilot or Prefect, inside the pipeline's budget.
Score checkpoints
Evaluate each checkpoint on held-out data against the limit you set.
Promote
A person sees the scores and decides. Nothing is promoted without that.
What the record keeps
The pinned dataset and base model, the training settings, the candidate adapter, checkpoint scores, the spend, and who approved the release.
Stop bad runs early
A dataset that fails its checks never reaches a GPU.
Prove it taught something
training/verifiable-finetune measures the base model first and re-measures after.
Close the loop to a callable model
training/dpo-train aligns the adapter on preferences; serving/serve puts the promoted adapter behind a vLLM endpoint and load-tests it.
Ships as training/e2e-finetune and training/verifiable-finetune. Engine support and recovery behavior vary by deployment.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run training/e2e-finetune