Model Fine-Tuning API

The Model Fine-Tuning API manages training jobs, reusable tuning configurations, checkpoints, lineage, deployment, and export.

View as Markdown Open the plain-text version of this page.

It is available only when the Node runtime provides the required stores, queues, and Python fine-tuning service.

Base Path#

/v1/projects/:projectId/fine-tuning

Reads require workspace:read or project:data:read. Mutations require models:manage, project:data:write, or project:models:manage according to the operation.

Jobs#

GET  /v1/projects/:projectId/fine-tuning/jobs
POST /v1/projects/:projectId/fine-tuning/jobs
GET  /v1/projects/:projectId/fine-tuning/jobs/:jobId
POST /v1/projects/:projectId/fine-tuning/jobs/:jobId/cancel

A job requires a known base model and a non-empty chat/v1 Dataset. Creation freezes the selected Examples into an immutable manifest and dispatches asynchronous training.

Supported values currently include:

engine: transformers
method: sft | dpo | orpo | kto
adapter: lora | qlora | full
output format: safetensors | gguf
GGUF quantization: q4_0 | q4_k_m | q4_k_s | q5_0 | q5_k_m | q6_k | q8_0 | f16 | bf16 | f32
Compressed quantization: fp8 | fp8_dynamic | fp8_static | int8

Checkpoints And Lineage#

GET  /v1/projects/:projectId/fine-tuning/checkpoints
GET  /v1/projects/:projectId/fine-tuning/checkpoints/:checkpointId
POST /v1/projects/:projectId/fine-tuning/checkpoints/:checkpointId/deploy
GET  /v1/projects/:projectId/fine-tuning/lineage/:rootModelId

Deployment requires a checkpoint with a registered inference profile and an available model runtime. The response is returned only after the service reaches running, fails, or times out.

Reusable Configurations#

GET    /v1/projects/:projectId/fine-tuning/tuning-configs
POST   /v1/projects/:projectId/fine-tuning/tuning-configs
GET    /v1/projects/:projectId/fine-tuning/tuning-configs/:configId
PATCH  /v1/projects/:projectId/fine-tuning/tuning-configs/:configId
DELETE /v1/projects/:projectId/fine-tuning/tuning-configs/:configId

Configurations bind a base model, Dataset, training settings, output settings, automatic trigger, and optional automatic deployment policy.

Export#

POST /v1/projects/:projectId/fine-tuning/export

Exports accept a checkpoint or Hugging Face model and publish to Hugging Face Hub. Supported export formats are gguf, merged-16bit, and lora-adapter. GGUF exports also require a quantization method.

See Model fine-tuning, Datasets, and Models API.