Finetuning Runs

A finetuning run trains one base model on one training dataset and, on success, registers a new model in your organization's registry.

Starting a Run

POST /v2/ai-platform/finetune-runs

{ "finetuneDatasetId": "ftd_abc123", "baseModelSlug": "qwen3.5-0.8b" }

The run is charged the base model's flat costCredits upfront (see Base Models) and created with status: "queued". Unlike every other queue-backed job in Crowdee, a finetuning run isn't dispatched over BullMQ — a separate Python worker (finetune-worker) polls finetune_jobs directly via Postgres and claims the oldest queued row.

Optional hyperparams (all have defaults):

{ "loraRank": 8, "loraAlpha": 16, "loraDropout": 0.05, "epochs": 3, "batchSize": 1, "learningRate": 0.0002 }

Checking Status

GET /v2/ai-platform/finetune-runs/{runId}

{
  "id": "ftj_abc123",
  "status": "running",
  "progress": 70,
  "baseModelSlug": "qwen3.5-0.8b",
  "resultModelArtifact": null
}

status is one of queued, running, completed, failed. progress is a 0–100 estimate updated by the worker as it moves through tokenization, training, and GGUF conversion. List every run for the active organization:

GET /v2/ai-platform/finetune-runs?status=completed

Model Registry

Once a run completes, resultModelArtifact points at a model_artifacts row:

GET /v2/ai-platform/models

[
  { "id": "ma_abc123", "name": "crowdee-org123-qwen3.5-0.8b", "baseModelSlug": "qwen3.5-0.8b", "ollamaTag": "crowdee-org123-qwen3.5-0.8b", "status": "ready" }
]

ollamaTag is the model's name in Crowdee's shared Ollama runtime — every organization's models are namespaced by organization ID, never shared across tenants.

Benchmarking a Model with the Crowd

There's no dedicated benchmarking endpoint — reuse the external-submission path of AI Result Evaluation: run your model against a validation set yourself, then submit its outputs via:

POST /v2/projects/{projectId}/ai-output-evaluations/external

This gets you back the same crowd-rated clarity/evidence/actionability/bias scores used to evaluate any other AI system's output — an independent signal before you rely on the model in production.

Roadmap

Not yet available, tracked for a future release:

• Closed-loop retraining — automatically kicking a new finetuning run off ai_component_feedback_log.flaggedForRetraining entries or Continuous Monitoring drift flags.
• Pipeline integration — using a finetuned model as a drop-in replacement for a Verification or Language Technology pipeline's built-in AI stage.
• Direct API/inference access to a finetuned model outside the platform.