Artifact types

Every link on the canvas carries one of eight artifact types; this page lists what each type means, which blocks produce and accept it, its colour on the canvas and how to convert between types.

Overview

A block's inputs and outputs are typed ports. You can only link an output to an input that accepts its type; the editor refuses other links and tells you which types the input expects. The types come from the block registry (tuneline/registry.py, registry version 1.0.0).

Type	Label in the app	Canvas colour (light / dark)	Meaning
Dataset	Dataset	ochre #AD6800 / #E3A24B	Rows loaded from a file, a URL, a golden set or the Hugging Face Hub, before formatting.
ChatDataset	Chat dataset	rose #BB3B72 / #F07FB0	Rows in chat form (messages) or prompt/completion form, ready for SFT or evaluation.
PreferenceDataset	Preference dataset	orange #C4501C / #F0895C	Prompt with chosen/rejected answers (DPO, ORPO, SimPO) or prompt/completion/label (KTO).
VisionDataset	Vision dataset	green #2F7D4F / #6FCA95	Chat rows with one or more images each, for vision-language models.
Model	Model	teal #0B7A90 / #4CC5DB	A base model, optionally with a LoRA adapter, or a merged fine-tuned model.
EvalReport	Eval report	violet #7650CB / #B499F3	Metrics and per-example results from an evaluation.
GGUF	GGUF files	brown #8A5A2B / #CF9A6B	Quantized GGUF files for llama.cpp, Ollama and LM Studio.
Endpoint	Endpoint	slate #4A5A56 / #9DB2AD	A running OpenAI-compatible inference endpoint.

The colours are the --t-* tokens in web/src/index.css: handles, link lines and type badges use them, so a rose handle always means chat data. The four dataset types (Dataset, ChatDataset, PreferenceDataset, VisionDataset) are grouped as dataset types in the registry.

Rules for links

• One link per input. Each input port accepts at most one link. Dropping a new link on a used input replaces the old one. An output can feed any number of inputs.
• Required and optional inputs. A required input must be linked before the pipeline can run; optional inputs (for example the eval input of Fine-tune (SFT)) can stay empty. The Block reference marks each input.
• Pass-through types (same_as). Some blocks work on any dataset type and give back the same type they received. Their outputs declare same_as: "<input>": the output takes the type of whatever is linked to that input. The editor resolves this from the upstream link when you connect the next block, and the validator does the same (resolve_output_types in tuneline/spec.py). If the input is not linked yet, the output's type is unknown until it is.

Block	Output	Takes the type of input
Clean and dedupe (data.clean)	data	data
Train / validation / test split (data.split)	train, validation, test	data
Merge datasets (data.merge)	data	a (link datasets of the same type to a and b)

So Load dataset → Format as chat → Split gives three ChatDataset outputs, and Load dataset → Split gives three plain Dataset outputs, which a trainer will not accept until they are formatted.

Conversions

When a link is refused, the editor suggests the block that converts between the two types:

From	To	Use
Dataset	ChatDataset	Format as chat (data.to_chat)
Dataset	PreferenceDataset	Format as preference pairs (data.to_preference)
Dataset	VisionDataset	Format as vision chat (data.to_vision)
Model	GGUF	Export GGUF (export.gguf)
Model	Endpoint	Deploy endpoint (deploy.endpoint)
GGUF	Endpoint	Ollama package (deploy.ollama)
Endpoint	EvalReport	Endpoint smoke test (deploy.smoke_test)

Two more blocks turn something into a Model without changing the type: Merge adapter (a base model plus adapter becomes one merged model) and Quality gate (passes the model on only when its rules hold). Generate synthetic data produces a ChatDataset from nothing or from an optional Dataset / ChatDataset seed.

Dataset

Rows loaded from a file, a URL, a golden set or the Hugging Face Hub, before formatting. Columns are whatever the source had; trainers do not accept this type directly.

Produced by	Accepted by
Load dataset data	Clean and dedupe data
Clean and dedupe data (pass-through)	Format as chat data
Train / validation / test split train, validation, test (pass-through)	Format as preference pairs data
Merge datasets data (pass-through)	Format as vision chat data
	Train / validation / test split data
	Merge datasets a, b
	Generate synthetic data seed (optional)

ChatDataset

Rows in chat form (a messages list of {role, content}) or prompt/completion form, ready for supervised fine-tuning or evaluation.

Produced by	Accepted by
Format as chat data	Fine-tune (SFT) train, eval (optional)
Generate synthetic data data	Reinforcement tuning (GRPO) train
Clean and dedupe, Split, Merge datasets (pass-through)	Evaluate on test set test
	Prompt-injection test test
	Clean and dedupe, Split, Merge datasets
	Generate synthetic data seed (optional)

PreferenceDataset

A prompt with a chosen and a rejected answer (for DPO, ORPO and SimPO), or a prompt, a completion and a good/bad label (for KTO).

Produced by	Accepted by
Format as preference pairs data	Preference tuning (DPO / ORPO / KTO) train, eval (optional)
Clean and dedupe, Split, Merge datasets (pass-through)	Clean and dedupe, Split, Merge datasets

VisionDataset

Chat rows with one or more images each, for vision-language models. Images can be Hub image columns, URLs, or files inside an uploaded ZIP.

Produced by	Accepted by
Format as vision chat data	Vision fine-tune (SFT) train, eval (optional)
Clean and dedupe, Split, Merge datasets (pass-through)	Evaluate on test set test
	Clean and dedupe, Split, Merge datasets

Model

A base model from the Hub, the same base model with a LoRA adapter, or a merged fine-tuned model. The artifact records the base model and its commit, the adapter or merged repository if any, the modality (text or vision) and the training method.

Produced by	Accepted by
Base model model	Fine-tune (SFT) model
Registered model model	Preference tuning model
Fine-tune (SFT) model	Vision fine-tune (SFT) model
Preference tuning model	Reinforcement tuning (GRPO) model
Vision fine-tune (SFT) model	Evaluate on test set model
Reinforcement tuning (GRPO) model	Benchmarks (lm-eval) model
Quality gate model (approved model)	Prompt-injection test model
Merge adapter model	Quality gate model
Publish and register model	Merge adapter model
	Export GGUF model
	Publish and register model
	Deploy endpoint model

Trainers take a Model and give back a Model, so you can chain them, for example SFT followed by DPO. Deploy endpoint needs the model on the Hugging Face Hub: keep Push to Hugging Face on in the trainer, or merge or publish it first.

EvalReport

Metrics and per-example results from an evaluation: a metrics object (for example label_accuracy, and with Also score the base model also base.<metric> and delta.<metric>), the number of examples n, and up to 50 sample rows with input, expected and actual output. Reports are stored inline with the run, not as files.

Produced by	Accepted by
Evaluate on test set report	Quality gate report, extra_report (optional)
Benchmarks (lm-eval) report	Publish and register report (optional: metrics go into the model card and registry)
Prompt-injection test report
Endpoint smoke test report

Gate rules refer to the metric names in these reports; see Evaluation and quality gates.

GGUF

Quantized GGUF files (for example q4_k_m and q8_0) made with llama.cpp, for Ollama, LM Studio and llama.cpp. The artifact lists one file per quantization and, when pushed, the Hugging Face repository that holds them.

Produced by	Accepted by
Export GGUF gguf	Ollama package gguf

GGUF files cannot be trained, evaluated or served by Deploy endpoint; deploy the adapter or merged model instead.

Endpoint

A running OpenAI-compatible inference endpoint: its URL (ending in /v1), the model name to send, and the provider (RunPod Serverless, Hugging Face Inference Endpoints, Modal, or Ollama on your own computer).

Produced by	Accepted by
Deploy endpoint endpoint	Endpoint smoke test endpoint
Ollama package endpoint

Endpoints created by a run also appear on the Deployments page; see Deployments.

Where artifacts are stored

Type	Stored as	Location
Dataset types	Parquet file	The machine that ran the step, and, when a Hugging Face token is saved, your private dataset repository <your-hf-username>/tuneline-cache at artifacts/<node key>/<port>.parquet
Model	Hub repository	Your Hugging Face account (adapter, merged model or published model), private by default; the base model stays where it is on the Hub
EvalReport	JSON	Inline in the run (artifacts.meta and the step's outputs)
GGUF	.gguf files	The machine that ran the step, and your Hugging Face account when Push the GGUF files is on
Endpoint	URL	At the provider; listed on the Deployments page

Artifact URIs use hf://model/<repo>@<revision>, hf://dataset/<repo>@<revision>/<path>, file://<path> (only valid on the machine that wrote it), inline: (data kept in the artifact record) and https://… (endpoints). See Architecture for how cached artifacts are found again.