Datasets

Datasets are the starting point of every Forge workflow. A dataset is a versioned container: every upload and every annotation export creates a new, immutable version, and every training, optimization or evaluation job references one exact version. That is what makes every model reproducible.

Dataset formats

Choose the format that matches the model family:

Format	Used for	Contents
COCO	CV detection / segmentation	Images + a COCO annotations.json
TXT	CV detection training	Images + one normalized detection-box .txt label per image + a manifest YAML
ImageNet	Image classification	Per-class folder layout
LLM-SFT	LLM / VLM fine-tuning	A sharegpt-style JSONL corpus; VLM datasets additionally carry an images/ folder
JSONL	Text corpora	One sample per line (plain text, {text}, {data} or sharegpt conversations)

llm-sft and imagenet versions are directly selectable for training. A raw COCO upload is converted for the trainer automatically (a manifest YAML is generated); TXT labels produced by an annotation export need no conversion at all.

A sharegpt line looks like this and feeds the LLM/VLM trainers directly:

{"conversations": [{"from": "human", "value": "Describe this image."},
                   {"from": "gpt", "value": "A blue circle on a red background."}]}

Create and upload

1. On the Datasets page, create a dataset with a name, description and format.
2. Upload individual images or a ZIP archive. Images may sit at the archive root or in an images/ subdirectory.
3. The archive is extracted and ingested asynchronously — follow progress in the version list. Each successful ingest becomes a new retained version.
4. Pre-staged corpora (for example an LLM-SFT JSONL plus an images/ directory already placed in object storage) can be registered from their S3 prefix without a browser upload.

Accepted image types are jpg / jpeg / png. Bulk bytes go to the S3-compatible object store; the workbench and trainers receive short-lived presigned URLs instead of file copies.

Versioning rules

• Every upload increments the version; old versions are retained and remain selectable. Versions are never mutated.
• An annotation export (COCO/TXT for vision, JSONL for text) is itself a new dataset version, so a model is always traceable to the exact labels it learned from.
• Switch between versions when creating a job; the rest of the pipeline does not change.
• The version records its file counts (images and annotation files), format, size and origin (upload, export, S3 reconciliation).

How versions feed the factory

upload images ──► v1 (raw)
annotate + export ──► v2 (TXT / JSONL, training-ready) ──► training
new field images ──► v3 ──► annotate gaps ──► v4 ──► retrain

FAQ

How do I prepare labeled training data?

Label the images in the annotation workbench, then use Export. Vision projects export COCO/TXT; text-classification exports plain JSONL; generation/chat projects export sharegpt JSONL. The result is a new dataset version the training page can select immediately.

How are version numbers incremented?

Automatically, once per finalized upload or export. You cannot overwrite a version.

Can a dataset mix images and text?

VLM corpora do exactly that: an llm-sft JSONL references files in an accompanying images/ directory. For pure vision use COCO/TXT/ImageNet; for pure text use JSONL/LLM-SFT without images.

Next steps

• Label the data on the Annotations page, or
• Start a Training run directly on a training-ready version.
• Set up pre-labeling first? See ML Backend.