Can an AI agent fine-tune a model for you, end to end?

A coding agent like Claude Code, Codex or Cursor can perform a full fine-tuning process from data validation to downloading the adapter. The platform imposes a monetary cap and you retain the responsibility for key decisions. On Nodus, the agent operates via the Nodus MCP server or CLI, suspending the TrainingJob when the $ maxCostUSD limit is reached. The agent manages launching, monitoring, evaluating, and terminating the run; handling data rights, defining improvement criteria, and shipping decisions are your domain.

The fine-tuning process mirrors the manual workflow with the agent generating commands while you intervene at three stages. To initiate, ensure the MCP setup is complete - the agent will guide you through this. Assess the provided JSONL data, verifying column mappings to prompt and completion fields, tallying the total rows, and flagging duplicates or empty answers.

Next, upload the data using the CLI command nodus volume put, ensuring the Volume status is Ready as the TrainingJob pins its latest revision. Draft the manifest outlining the TrainingJob on the nodus/sft runtime, incorporating a LoRA block, validation split, and the $ maxCostUSD cap. Utilize the MCP estimate tool or nodus apply -f sft.yaml --dry-run -o yaml to ascertain completion cost and any potential blockers.

Upon approval, launch the job. Over MCP, the first apply returns a dry run and etag; the agent repeats the apply command with confirmed: true and identical etag. Monitor the run's progress - the agent polls with wait, describe, and log commands or nodus get tj -w on the CLI, displaying stage, step, loss, and cost per stage. Evaluate the outputs/results.json file and download the adapter using nodus cp, verifying each file's sha256 checksum.

Any parameter name typos, such as learning_rate, will be rejected with the field path before any charges occur.

The manifest structure adheres to the documented LoRA example, scaled for an 8B model on one GPU. The manifest specifies model name, step count, cost cap, runtime, model URI, data volume and format, validation fraction, parameters, and GPU resources. It explicitly defines the maxCostUSD to $15.00, providing a safety net against unexpected overruns.

The cost comprises GPU-hours multiplied by the listed rate, with the cap acting as a ceiling. For instance, the dry run estimates 4 hours, translating to $10.40 at $2.60/hr for the H100 SXM GPU. Setting the cap at $15.00 allows for a 200-step smoke test costing $2.00, ensuring any data or config errors are caught early. The maximum potential cost remains $17.00, even if unforeseen issues arise.

In cases where the run exceeds the $15.00 cap, it is suspended rather than failed, allowing a resume from the checkpoint if the cap is increased. Various GPU options are available, each with distinct hourly rates ($0.793 for L40S, $1.090 for A100, and $1.350 for RTX PRO 6000). Specify the resources section in the manifest to select an appropriate GPU, or let the runtime automatically choose the most suitable one. Dry-run estimates guide you in selecting the most cost-effective GPU.