Glossary

Short definitions of the terms used in Tuneline Studio and in these docs, in alphabetical order.

Adapter

A small set of extra weights trained on top of a frozen base model with LoRA or QLoRA. Adapters are a few megabytes to a few hundred megabytes and are pushed to your Hugging Face account as <pipeline>-adapter by default. To serve or share the model as one piece, merge the adapter into the base model (merged model). In the Registry, adapter versions have kind adapter.

Artifact

Anything a step produces and passes on through an output: a dataset, a model, an evaluation report, GGUF files or an endpoint. Each has one of eight artifact types and a URI such as hf://model/<user>/<repo>@<revision>. The Artifacts tab of a run lists them.

Base model

The pretrained model you start from, picked in the Base model block from the catalog or any Hugging Face repository id (for example Qwen/Qwen3-4B-Instruct-2507). Evaluation can also score the base model (Also score the base model) so you see the gain from fine-tuning.

Block

Also called step or node. One box on the canvas, such as Load dataset or Fine-tune (SFT). Each block has a type id (train.sft), typed inputs and outputs, and settings. In a run, each block becomes a step with its own status, logs and outputs. See the Block reference.

Cache key

A fingerprint of a step: its block type and version, the settings that affect the result, the cache keys of its inputs and the identity of its data (file checksum, Hub commit). Two steps with the same cache key produce the same result.

Cached step

A step that was not executed because a previous run already produced every one of its outputs with the same cache key. It shows as cached on the run page and its outputs are reused. Turn off Reuse cached steps in the Run dialog to recompute everything. Blocks that act on the outside world (publish, deploy, gate) are never cached.

ChatDataset

The artifact type for rows in chat form (messages with roles) or prompt/completion form, ready for SFT or evaluation. Make one from a plain Dataset with Format as chat.

DPO

DPO stands for Direct Preference Optimization. A preference-tuning method: from pairs of a preferred (chosen) and a rejected answer to the same prompt, it teaches the model to prefer the chosen one, using the starting model as a reference. The default algorithm of Preference tuning; usually run after SFT.

Endpoint

A running, OpenAI-compatible inference server for your model (vLLM on RunPod Serverless, Hugging Face Inference Endpoints or Modal, or Ollama on your own computer). Created by the Deploy endpoint block or on the Deployments page; you can chat with it in the Playground.

Epoch

One full pass over the training data. Epochs can be fractional (for example 0.5); Max steps overrides it when set.

Estimate

Tuneline's prediction of the GPU memory a training step needs, which GPUs it fits on, and how long and how much it will take on each provider. Shown on the Hardware page and in the Run dialog. It is an estimate: the assumptions are listed with it.

Full fine-tuning

Training every weight of the model (method full) instead of an adapter. It needs roughly 16 to 18 bytes of GPU memory per parameter, so it suits small models only.

Gate

The Quality gate block. It passes the model on only if every rule holds, for example label_accuracy >= 0.9, using metrics from evaluation reports (base. and delta. prefixes compare with the base model). A failed gate fails the run, or with Skip later steps skips what comes after it. See Evaluation and quality gates.

GGUF

The single-file model format used by llama.cpp, Ollama and LM Studio. The Export GGUF block converts a model to GGUF in one or more quantizations.

Golden set

A set of hand-picked example conversations with expected answers, saved from the Playground (Save to golden set). Load dataset can read a golden set as test data so every new model version is scored on the same examples.

GRPO

GRPO stands for Group Relative Policy Optimization. A reinforcement-learning method: the model writes several answers per prompt, reward functions score them (exact or contained answer, format regex, valid JSON, length penalty), and the model is pushed towards the better ones. Used by Reinforcement tuning (GRPO) for reasoning tasks with checkable answers.

KTO

KTO stands for Kahneman-Tversky Optimization. A preference-tuning method that learns from single answers labelled good or bad instead of chosen/rejected pairs. Choose kto in Preference tuning, with unpaired data from Format as preference pairs.

LoRA

LoRA stands for Low-Rank Adaptation. Fine-tuning that freezes the base model and trains small low-rank matrices added to its layers (an adapter). The LoRA rank (lora_r) sets the adapter's size; 8 to 32 covers most tasks.

Merged model

A base model with an adapter folded into its weights, made by the Merge adapter block. It loads like any model without the adapter library and is what Export GGUF converts. In the Registry, merged versions have kind merged.

MFU

MFU stands for model FLOPs utilisation. The share of a GPU's peak compute that training actually uses. The estimator assumes 35% by default to turn the work of a training run into hours.

ORPO

ORPO stands for Odds Ratio Preference Optimization. A preference-tuning method that combines supervised fine-tuning and preference learning in one step and needs no reference model. Choose orpo in Preference tuning.

Pipeline

The graph of blocks and links you build on the canvas: where data comes from, how the model is trained, evaluated, exported and deployed. Pipelines are saved automatically and can be exported as YAML.

QLoRA

LoRA on a base model loaded in 4-bit precision, which cuts memory several times over. The default method of the training blocks and the usual choice for 7B+ models on one GPU. It needs an NVIDIA GPU with bitsandbytes (most reliable on Linux or WSL2).

Quantization

Storing weights with fewer bits to make a model smaller and faster to run. GGUF quantizations offered by Export GGUF are f16, bf16, q8_0, q6_k, q5_k_m, q4_k_m, q3_k_m and q2_k: the number is roughly the bits per weight, so q8_0 is close to the original quality and q4_k_m is a common balance of size and quality. QLoRA uses a different, 4-bit quantization during training.

Registry

Two things in Tuneline. The model registry (Registry page) lists your models and their versions with metrics, lineage and stage (staging, production, archived); Publish and register adds versions. The block registry (tuneline/registry.py) defines every block type, its ports and settings; the Block reference is generated from it.

Run

One execution of a pipeline on a chosen target. A run has a number, a status (queued, launching, running, succeeded, failed, cancelled), a step status per block, live logs and metrics, artifacts, and its cost. See Monitoring runs.

Run token

A secret starting with tlr_ that lets a runner fetch one run and report its progress. It gives access to that run only and to the API keys its pipeline needs. A run token expires after 14 days (30 at most), and in any case 2 hours after the run ends. Several tokens can be valid for the same run. They appear in the commands and notebooks the app gives you, so keep those private.

Scale to zero

An endpoint setting where no GPU runs, and nothing is billed, while nobody sends requests. RunPod Serverless and Hugging Face Inference Endpoints scale to zero after the idle time you set (Scale down after idle); Modal stops after idle. The first request after a pause waits for a cold start while the model loads.

SFT

SFT stands for supervised fine-tuning. Training a model on example conversations so it learns to answer like them. The Fine-tune (SFT) and Vision fine-tune (SFT) blocks do this with LoRA, QLoRA or full fine-tuning.

SimPO

SimPO stands for Simple Preference Optimization. A reference-free variant of DPO: it learns from chosen/rejected pairs without keeping a copy of the starting model. Choose simpo in Preference tuning.

Target

Where a run executes: modal, runpod, hf_jobs (Hugging Face Jobs), kaggle, notebook (Colab, Kaggle, Lightning or any Jupyter) or local (My computer). You choose it in the Run dialog.

Template

A ready-made pipeline to start from, such as Sanctions alert triage or CPU smoke test. See Templates.

vLLM

An open-source inference engine for LLMs. Tuneline's cloud endpoints serve models with vLLM behind an OpenAI-compatible API, and Evaluate on test set can use it to generate answers faster on NVIDIA GPUs.

Worker

The tuneline worker process on your own computer. It checks for runs you send to My computer, runs them and reports back, and shows the computer as online under Settings → Your computers. See Runner CLI.

Worker token

A secret starting with tlw_ that connects one computer to your account. It is shown once when you add the computer and works until you remove that computer in Settings.