Training Data
How a finetuning dataset is assembled from a Data Platform dataset version or a Crowd Platform job's accepted answers.
Every finetuning run needs a finetune dataset — a training-pairs manifest built from one of two sources. The AI Platform never collects labels itself; it turns data you already have on the Data Platform or Crowd Platform into a ready-to-use training set.
Source 1: A Data Platform Dataset Version
If a dataset version has already been through the Crowd-Assisted Labeling pipeline (crowd-label-classify), the AI Platform reads its per-file labels (data_platform_file_labels, the majority-vote result of that pipeline's crowd consensus) and pairs each file with its label:
{ "fileId": "file_abc123", "fileName": "clip-01.txt", "mimeType": "text/plain", "fileUrl": "https://...", "label": "spam" }A dataset version with no completed labeling run has nothing to build a classification training set from — build one with crowd-label-classify first.
v1 finetuning is text-only. Non-text files (image/audio/video) can still be part of a labeled dataset version, but rows referencing them are skipped when the finetune-worker builds the actual training text. This applies uniformly across the whole catalog: some base models (Ministral 3, Gemma 4, Qwen3.5) are natively multimodal checkpoints, but the finetune-worker only ever trains their language-model component on text pairs — never images or audio — so the data you provide is the same regardless of which base model you pick. See Base Models for which models are multimodal-native.
Source 2: A Crowd Platform Job's Accepted Answers
Any crowdsourcing job's accepted answers can become instruction-following training data. Each row pairs the worker's task input snapshot with their answer:
{ "input": { "prompt": "..." }, "output": { "category": "..." } }This is useful for tasks where your crowd has already produced the ground truth you want a model to learn to reproduce — for example, distilling a recurring crowd task into a cheap model that pre-filters or automates it going forward.
Don't have a job like this yet? See Instruction/Response Template for a pre-built SurveyJS template and step-by-step guide to setting one up from scratch.
Building the Dataset
POST /v2/ai-platform/finetune-datasets{ "sourceType": "dataset_version", "datasetId": "dataset_abc123", "versionId": "version_def456" }or
{ "sourceType": "crowd_job", "jobId": "job_abc123" }The response is a finetune_datasets row: an id, format (classification or instruction), and rowCount. The underlying manifest is stored as JSONL in S3 — the finetune-worker resolves file content and builds the final training text at train time, so this step stays fast regardless of file size.
Next Step
Pick a base model and start a finetuning run against the assembled dataset.
How is this guide?