Block reference

Every block you can place on the canvas (29 in registry version 1.0.0), with its inputs, outputs, settings, presets and the API keys it needs.

Note

Generated from the block registry by documentation/scripts/gen_reference.py — don't edit by hand. To change this page, change tuneline/registry.py or tuneline/templates.py, regenerate web/src/generated/*.json, then run the script again.

How to read this page

• Type is the id used in pipeline YAML (type: train.sft) and in run specs; version is the block's version, which is part of the cache key.
• Runtime says what the block needs: cpu, gpu or network (internet only). A pipeline needs a GPU when any of its blocks has runtime gpu, unless you choose a CPU-only run for a tiny model.
• Cached blocks can be skipped on a later run when the same inputs and settings were already computed (see Running pipelines). Blocks that always run act on the outside world (publish, deploy, gate).
• Inputs marked required must be linked before the pipeline can run; each input accepts one link. The types link to Artifact types.
• Settings are grouped as in the inspector. Shown when is the condition under which the inspector shows a setting (and the validator checks it); hidden settings keep their default. Setting keys are what you write under with: in pipeline YAML and in tuneline run-local --set NODE.KEY=VALUE.
• API keys are saved under Settings → API keys (see API keys and security). Required keys block the run until they are saved; optional keys are used when present.

Setting kinds:

Kind	Meaning
text	One line of text.
textarea	Several lines of text (templates, prompts).
number	A decimal number.
integer	A whole number.
select	One value from a fixed list of choices.
multiselect	Any number of values from a fixed list (a list in YAML).
boolean	On or off (true / false).
model	A Hugging Face model id such as Qwen/Qwen3-4B-Instruct-2507, picked with the model picker.
dataset_version	The id of one of your dataset versions (Datasets page).
eval_set	The id of one of your golden sets (Playground).
model_version	The id of a model version in your Registry.
rules	Quality-gate rules, one per line, such as label_accuracy >= 0.9.
json	A JSON value.

Text settings can contain {pipeline} (the pipeline's slug), {run} (the run number) and {app_url} (the app's address); the runner fills them in.

Contents

Category	Block	Type	Takes	Gives	Runtime
Data	Load dataset	data.load	—	Dataset	network
Data	Clean and dedupe	data.clean	ChatDataset, Dataset, PreferenceDataset, VisionDataset	ChatDataset, Dataset, PreferenceDataset, VisionDataset	cpu
Data	Format as chat	data.to_chat	Dataset	ChatDataset	cpu
Data	Format as preference pairs	data.to_preference	Dataset	PreferenceDataset	cpu
Data	Format as vision chat	data.to_vision	Dataset	VisionDataset	network
Data	Train / validation / test split	data.split	ChatDataset, Dataset, PreferenceDataset, VisionDataset	ChatDataset, Dataset, PreferenceDataset, VisionDataset	cpu
Data	Merge datasets	data.merge	ChatDataset, Dataset, PreferenceDataset, VisionDataset	ChatDataset, Dataset, PreferenceDataset, VisionDataset	cpu
Data	Generate synthetic data	data.synthetic	ChatDataset, Dataset	ChatDataset	network
Data	Image data (folder)	data.images	—	ImageData	cpu
Model	Base model	model.base	—	Model	network
Model	Registered model	model.registry	—	Model	network
Train	Fine-tune (SFT)	train.sft	ChatDataset, Model	Model	gpu
Train	Preference tuning (DPO / ORPO / KTO)	train.preference	Model, PreferenceDataset	Model	gpu
Train	Vision fine-tune (SFT)	train.vision_sft	Model, VisionDataset	Model	gpu
Train	Reinforcement tuning (GRPO)	train.grpo	ChatDataset, Model	Model	gpu
Train	Train a segmenter (U-Net)	train.segmenter	ImageData	EvalReport, VisionNet	gpu
Train	Train a text reader (CRNN)	train.text_reader	ImageData	EvalReport, VisionNet	gpu
Train	Train an image embedder (similarity)	train.embedder	ImageData	EvalReport, VisionNet	gpu
Evaluate	Evaluate on test set	eval.golden	ChatDataset, Model, VisionDataset	EvalReport	gpu
Evaluate	Benchmarks (lm-eval)	eval.benchmarks	Model	EvalReport	gpu
Evaluate	Prompt-injection test	eval.redteam	ChatDataset, Model	EvalReport	gpu
Evaluate	Quality gate	eval.gate	EvalReport, Model, VisionNet	Model, VisionNet	cpu
Export	Merge adapter	export.merge	Model	Model	gpu
Export	Export GGUF	export.gguf	Model	GGUF	cpu
Export	Publish and register	export.hub	EvalReport, Model	Model	network
Export	Export to ONNX	export.onnx	VisionNet	VisionNet	cpu
Deploy	Deploy endpoint	deploy.endpoint	Model	Endpoint	network
Deploy	Ollama package	deploy.ollama	GGUF	Endpoint	cpu
Deploy	Endpoint smoke test	deploy.smoke_test	Endpoint	EvalReport	network

Data

Load dataset

Type data.load · version 1.0 · runtime: network (needs internet access only) · results are cached

Reads rows from an uploaded file, the Hugging Face Hub, a URL or one of your golden sets.

Inputs

None: this block starts a branch of the pipeline.

Outputs

Output	Label	Type
data	Rows	Dataset

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
source	Source (required)	select	upload	upload (Uploaded file) hf (Hugging Face dataset) url (URL (CSV, JSONL, JSON, Parquet or ZIP)) eval_set (Golden set)	Always	—
dataset_version_id	Dataset (required)	dataset_version	—	—	source is upload	—
hf_repo	Dataset repository (required)	text	(empty)	—	source is hf	Example: HuggingFaceH4/no_robots
hf_config	Config / subset	text	(empty)	—	source is hf	—
hf_split	Split	text	train	—	source is hf	—
url	File URL (required)	text	(empty)	—	source is url	—
eval_set_id	Golden set (required)	eval_set	—	—	source is eval_set	—

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
hf_revision	Revision	text	(empty)	—	source is hf	Pin a commit for reproducible runs.
file_format	File format	select	auto	auto jsonl json csv parquet zip	Always	—
max_rows	Max rows (0 = all)	integer	0	0 to 100000000	Always	—
shuffle_seed	Shuffle before max rows (-1 = no shuffle)	integer	-1	-1 to 2147483647	Always	—

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	optional	When source is hf	Private or gated Hugging Face datasets

Clean and dedupe

Type data.clean · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Drops empty and duplicate rows, filters by length and redacts personal data such as emails, phone numbers, IBANs and card numbers.

Inputs

Input	Label	Accepts	Link
data	Rows	Dataset, ChatDataset, PreferenceDataset, VisionDataset	required

Outputs

Output	Label	Type
data	Clean rows	Same type as the input data (one of Dataset, ChatDataset, PreferenceDataset, VisionDataset)

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
drop_empty	Drop empty rows	boolean	true	—	Always	—
dedupe	Remove duplicates	select	exact	none (Keep duplicates) exact (Exact duplicates)	Always	—
dedupe_on	Compare columns (blank = all)	text	(empty)	—	Always	Example: prompt,response
min_chars	Min characters per row (0 = off)	integer	0	0 to 10000000	Always	—
max_chars	Max characters per row (0 = off)	integer	0	0 to 100000000	Always	—
pii_redact	Redact personal data	boolean	false	—	Always	—
pii_types	What to redact	multiselect	["email", "phone", "iban", "card", "ip"]	email phone iban card ip ssn url	pii_redact is true	—
text_columns	Columns to redact (blank = all text)	text	(empty)	—	pii_redact is true	—

API keys

None.

Format as chat

Type data.to_chat · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Turns rows into chat examples. Detects messages, ShareGPT, Alpaca and prompt/completion formats, or builds prompts from your own columns with templates.

Inputs

Input	Label	Accepts	Link
data	Rows	Dataset	required

Outputs

Output	Label	Type
data	Chat rows	ChatDataset

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
input_format	Input format	select	auto	auto (Detect automatically) messages (OpenAI messages) sharegpt (ShareGPT (conversations: from/value)) alpaca (Alpaca (instruction/input/output)) prompt_completion (Prompt and completion columns) columns (Build from columns with templates) text (Plain text (continued pretraining))	Always	—
messages_column	Messages column	text	messages	—	input_format is messages or sharegpt	—
prompt_column	Prompt column	text	prompt	—	input_format is prompt_completion	—
completion_column	Completion column	text	completion	—	input_format is prompt_completion	—
prompt_template	User message template	textarea	{input}	—	input_format is columns	Python format string over the row's columns, e.g. 'Alert: {alert_text}\nCustomer: {customer}'.
response_template	Assistant message template	textarea	{output}	—	input_format is columns	—
text_column	Text column	text	text	—	input_format is text	—
system_prompt	System prompt (optional)	textarea	(empty)	—	Always	Added to every example that has no system message.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
output_style	Output style	select	messages	messages (Chat messages) prompt_completion (Prompt / completion)	Always	—
keep_columns	Keep extra columns (comma separated)	text	(empty)	—	Always	For example a label column used later by evaluation.

Choices for output_style:

• messages — Chat messages: Uses the model's chat template (recommended).
• prompt_completion — Prompt / completion

API keys

None.

Format as preference pairs

Type data.to_preference · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Builds preference data: a prompt with a chosen and a rejected answer (DPO, ORPO, SimPO) or prompt, completion and a good/bad label (KTO).

Inputs

Input	Label	Accepts	Link
data	Rows	Dataset	required

Outputs

Output	Label	Type
data	Preference rows	PreferenceDataset

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
input_format	Input format	select	auto	auto (Detect automatically) paired (prompt / chosen / rejected) unpaired (prompt / completion / label (KTO))	Always	—
prompt_column	Prompt column	text	prompt	—	Always	—
chosen_column	Chosen column	text	chosen	—	input_format is auto or paired	—
rejected_column	Rejected column	text	rejected	—	input_format is auto or paired	—
completion_column	Completion column	text	completion	—	input_format is unpaired	—
label_column	Label column (true = good)	text	label	—	input_format is unpaired	—
system_prompt	System prompt (optional)	textarea	(empty)	—	Always	—

API keys

None.

Format as vision chat

Type data.to_vision · version 1.0 · runtime: network (needs internet access only) · results are cached

Builds image + text chat examples for vision-language models. Images can be Hub image columns, URLs, or files inside an uploaded ZIP.

Inputs

Input	Label	Accepts	Link
data	Rows	Dataset	required

Outputs

Output	Label	Type
data	Vision rows	VisionDataset

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
input_format	Input format	select	auto	auto (Detect automatically) messages (Messages with image parts) columns (Image column + templates)	Always	—
image_column	Image column	text	image	—	Always	A Hub image feature, an http(s) URL, or a file name inside the uploaded ZIP.
prompt_template	User message template	textarea	{question}	—	input_format is columns	—
response_template	Assistant message template	textarea	{answer}	—	input_format is columns	—
system_prompt	System prompt (optional)	textarea	(empty)	—	Always	—

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
max_image_size	Resize longest side to (px, 0 = keep)	integer	1024	0 to 8192	Always	—

API keys

None.

Train / validation / test split

Type data.split · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Splits rows into training, validation and test sets, optionally stratified by a label column.

Inputs

Input	Label	Accepts	Link
data	Rows	Dataset, ChatDataset, PreferenceDataset, VisionDataset	required

Outputs

Output	Label	Type
train	Train	Same type as the input data (one of Dataset, ChatDataset, PreferenceDataset, VisionDataset)
validation	Validation	Same type as the input data (one of Dataset, ChatDataset, PreferenceDataset, VisionDataset)
test	Test	Same type as the input data (one of Dataset, ChatDataset, PreferenceDataset, VisionDataset)

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
test_size	Test share	number	0.1	0 to 0.9, step 0.01	Always	—
validation_size	Validation share	number	0.05	0 to 0.9, step 0.01	Always	0 leaves the validation output empty.
stratify_column	Stratify by column (optional)	text	(empty)	—	Always	—
seed	Seed	integer	42	0 to 2147483647	Always	—

API keys

None.

Merge datasets

Type data.merge · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Concatenates two datasets of the same type, for example real and synthetic examples.

Inputs

Input	Label	Accepts	Link
a	First	Dataset, ChatDataset, PreferenceDataset, VisionDataset	required
b	Second	Dataset, ChatDataset, PreferenceDataset, VisionDataset	required

Outputs

Output	Label	Type
data	Merged	Same type as the input a (one of Dataset, ChatDataset, PreferenceDataset, VisionDataset)

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
shuffle	Shuffle after merging	boolean	true	—	Always	—
seed	Seed	integer	42	0 to 2147483647	shuffle is true	—

API keys

None.

Generate synthetic data

Type data.synthetic · version 1.0 · runtime: network (needs internet access only) · results are cached

Asks a strong model to write new training examples from your instructions and optional seed examples. Uses Hugging Face Inference Providers or any OpenAI-compatible API.

Inputs

Input	Label	Accepts	Link
seed	Seed examples	Dataset, ChatDataset	optional

Outputs

Output	Label	Type
data	Generated chat rows	ChatDataset

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
provider	Generator API	select	hf_router	hf_router (Hugging Face Inference Providers (uses your HF token)) openai_compatible (OpenAI-compatible API (uses your saved key and base URL))	Always	—
model	Generator model	text	Qwen/Qwen3-235B-A22B-Instruct-2507	—	Always	Any chat model your provider serves.
instructions	What to generate (required)	textarea	Write a realistic user request and an ideal assistant answer for the task described below.\n\nTask:	—	Always	—
num_examples	Number of examples	integer	100	1 to 20000	Always	—
few_shot	Seed examples shown per request	integer	3	0 to 20	Always	—
temperature	Temperature	number	0.8	0 to 2, step 0.05	Always	—
max_tokens	Max tokens per example	integer	1024	64 to 16384	Always	—
concurrency	Parallel requests	integer	4	1 to 32	Always	Not part of the cache key: changing it does not re-run a cached step.
system_prompt	System prompt for the trained model (optional)	textarea	(empty)	—	Always	—

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	When provider is hf_router	Hugging Face Inference Providers
openai_compat_key	API key for an OpenAI-compatible endpoint (judge / synthetic data)	required	When provider is openai_compatible	OpenAI-compatible API
openai_compat_base_url	Base URL for that OpenAI-compatible endpoint	required	When provider is openai_compatible	OpenAI-compatible API

Image data (folder)

Type data.images · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Points at labelled images in a folder on the computer that runs the job: segmentation arrays (images + per-pixel class maps) or image shards with text and group labels.

Inputs

None: this block starts a branch of the pipeline.

Outputs

Output	Label	Type
data	Image data	ImageData

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
path	Folder (required)	text	(empty)	—	Always	Path on the machine that runs the job (inside the container for Docker workers). Example: /data/images
format	Format (required)	select	image_shards	seg_cache (Segmentation arrays) image_shards (Image shards)	Always	—
val_split	Validation split name	text	test	—	format is seg_cache	Images whose meta.json split equals this are used for validation.
val_shards	Validation shards (last N files)	integer	2	0 to 10000	format is image_shards	—

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
pattern	File pattern	text	*.pkl	—	format is image_shards	—

Choices for format:

• seg_cache — Segmentation arrays: x.npy images, y.npy class maps, optional m.npy binary masks and meta.json with a split per image.
• image_shards — Image shards: *.pkl files of rows {png, text?, group?, negative_of?, tag?}.

API keys

None.

Model

Base model

Type model.base · version 1.0 · runtime: network (needs internet access only) · results are cached

Any open-weight model on the Hugging Face Hub that Transformers can load, text or vision. Optionally start from an existing LoRA adapter.

Inputs

None: this block starts a branch of the pipeline.

Outputs

Output	Label	Type
model	Model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
repo_id	Model (required)	model	Qwen/Qwen3-4B-Instruct-2507	—	Always	Pick from the catalog or paste any Hub repository id.
modality	Modality	select	auto	auto (Detect from the model config) text (Text only) vision (Vision + text)	Always	—

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
revision	Revision	text	(empty)	—	Always	Branch, tag or commit. Blank = main.
adapter_repo	Start from LoRA adapter (optional)	text	(empty)	—	Always	Example: your-name/previous-adapter
trust_remote_code	Trust remote code	boolean	false	—	Always	Some models ship custom Python code. Only enable for repositories you trust.
attn_implementation	Attention implementation	select	auto	auto sdpa flash_attention_2 eager	Always	Not part of the cache key: changing it does not re-run a cached step.

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	optional	Always	Gated models such as Llama and Gemma 2/3

Registered model

Type model.registry · version 1.0 · runtime: network (needs internet access only) · results are cached

A model version from your registry, for further training, evaluation or deployment.

Inputs

None: this block starts a branch of the pipeline.

Outputs

Output	Label	Type
model	Model	Model

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
model_version_id	Model version (required)	model_version	—	—	Always	—

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	Always	Your registered models live in your Hugging Face account

Train

Fine-tune (SFT)

Type train.sft · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Supervised fine-tuning with TRL: LoRA, QLoRA or full fine-tuning, with checkpoints, early stopping and live loss charts.

Inputs

Input	Label	Accepts	Link
train	Training data	ChatDataset	required
model	Base model	Model	required
eval	Validation data	ChatDataset	optional

Outputs

Output	Label	Type
model	Fine-tuned model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
method	Method	select	qlora	qlora (QLoRA (4-bit base + LoRA)) lora (LoRA (16-bit base + LoRA)) full (Full fine-tune)	Always	How much of the model is trained.
epochs	Epochs	number	1	0.1 to 50, step 0.1	Always	—
max_steps	Max steps (-1 = use epochs)	integer	-1	-1 to 1000000, step 1	Always	Stop after this many optimizer steps. Handy for quick tests.
learning_rate	Learning rate	number	0.0002	1e-08 to 0.1, step 1e-06	Always	—
max_seq_length	Max sequence length (tokens)	integer	2048	64 to 262144, step 64	Always	Longer examples are truncated. Memory grows with this.

LoRA adapter

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lora_r	LoRA rank (r)	integer	16	1 to 512, step 1	method is lora or qlora	Adapter size. 8-32 covers most tasks. vLLM serves ranks up to its --max-lora-rank (default 16).
lora_alpha	LoRA alpha	integer	32	1 to 1024, step 1	method is lora or qlora	Scaling factor. Common choice: 2 x rank.
lora_dropout	LoRA dropout	number	0.0	0 to 0.5, step 0.01	method is lora or qlora	—
target_modules	Target layers	select	all-linear	all-linear (All linear layers) attention (Attention only (q, k, v, o)) custom (Custom list)	method is lora or qlora	—
custom_target_modules	Custom target modules	text	(empty)	—	target_modules is custom	Example: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
use_dora	Use DoRA	boolean	false	—	method is lora or qlora	Weight-decomposed LoRA. Sometimes better quality, slower.
use_rslora	Use rank-stabilized LoRA	boolean	false	—	method is lora or qlora	Scales by alpha/sqrt(r); helps at high ranks.

Batch and memory

Setting	Label	Kind	Default	Choices / range	Shown when	Help
per_device_train_batch_size	Batch size per GPU	integer	2	1 to 512, step 1	Always	—
gradient_accumulation_steps	Gradient accumulation	integer	8	1 to 512, step 1	Always	Effective batch = batch size x accumulation x GPUs.
gradient_checkpointing	Gradient checkpointing	boolean	true	—	Always	Saves a lot of memory for ~20-30% more time.
completion_only_loss	Train on responses only	boolean	true	—	Always	Ignore the prompt tokens in the loss. Usually better for chat data.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lr_scheduler_type	LR schedule	select	cosine	cosine linear constant constant_with_warmup	Always	—
warmup_ratio	Warmup ratio	number	0.05	0 to 0.5, step 0.01	Always	—
weight_decay	Weight decay	number	0.01	0 to 1, step 0.001	Always	—
max_grad_norm	Max gradient norm	number	1.0	0 to 100, step 0.1	Always	—
optim	Optimizer	select	auto	auto (Auto) adamw_torch_fused (AdamW (fused)) adamw_torch (AdamW) adamw_8bit (AdamW 8-bit) paged_adamw_8bit (Paged AdamW 8-bit) adafactor (Adafactor)	Always	—
precision	Precision	select	auto	auto (Auto) bf16 fp16 fp32	Always	—
seed	Seed	integer	42	0 to 2147483647, step 1	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
logging_steps	Log every N steps	integer	5	1 to 10000	Always	Not part of the cache key: changing it does not re-run a cached step.
eval_strategy	Evaluate during training	select	epoch	no steps epoch	Always	Only used when the eval input is connected.
eval_steps	Eval every N steps	integer	100	1 to 1000000	eval_strategy is steps	—
save_steps	Checkpoint every N steps	integer	200	1 to 1000000	Always	Checkpoints let a stopped or preempted job resume. Not part of the cache key: changing it does not re-run a cached step.
early_stopping_patience	Early stopping patience (0 = off)	integer	0	0 to 100	Always	Stop when eval loss fails to improve this many evaluations in a row.
report_to	Also log to	select	none	none (Tuneline only) wandb (Weights & Biases)	Always	Not part of the cache key: changing it does not re-run a cached step.

Speed-ups

Setting	Label	Kind	Default	Choices / range	Shown when	Help
use_unsloth	Use Unsloth	boolean	false	—	Always	Faster, lower-memory training on one NVIDIA GPU for supported models. Current Unsloth releases need an older TRL than Tuneline uses, so until Unsloth supports TRL 1 this falls back to standard TRL (logged in the run).
use_liger_kernel	Use Liger kernels	boolean	false	—	Always	Fused kernels that cut memory for supported architectures.
packing	Pack short examples	boolean	false	—	Always	Concatenates examples to fill each sequence. Faster on short data.

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the adapter to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-adapter	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
neftune_noise_alpha	NEFTune noise (0 = off)	number	0.0	0 to 20, step 0.5	Always	—

Choices for method:

• qlora — QLoRA (4-bit base + LoRA): Lowest memory. Good default for 7B+ models on one GPU.
• lora — LoRA (16-bit base + LoRA): Slightly better quality, about 2-3x the memory of QLoRA.
• full — Full fine-tune: Updates every weight. Needs ~16-18 bytes per parameter; small models only.

Choices for target_modules:

• all-linear — All linear layers: Recommended (attention + MLP).
• attention — Attention only (q, k, v, o)
• custom — Custom list

Choices for optim:

• auto — Auto: paged 8-bit AdamW for QLoRA, fused AdamW otherwise
• adamw_torch_fused — AdamW (fused)
• adamw_torch — AdamW
• adamw_8bit — AdamW 8-bit
• paged_adamw_8bit — Paged AdamW 8-bit
• adafactor — Adafactor

Choices for precision:

• auto — Auto: bf16 on Ampere or newer, fp16 on T4/V100, fp32 on CPU
• bf16
• fp16
• fp32

API keys

Key	What it is	Need	When	Why
wandb_api_key	Weights & Biases API key	required	When report_to is wandb	Log training to Weights & Biases
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the adapter

Presets

Choosing a preset in the inspector replaces these settings; the rest keep their values.

Preset	Values
QLoRA quick	method = qlora lora_r = 16 lora_alpha = 32 lora_dropout = 0.0 learning_rate = 0.0002 epochs = 1 per_device_train_batch_size = 2 gradient_accumulation_steps = 8 max_seq_length = 2048 lr_scheduler_type = cosine warmup_ratio = 0.05
LoRA quality	method = lora lora_r = 32 lora_alpha = 64 lora_dropout = 0.05 learning_rate = 0.0001 epochs = 2 per_device_train_batch_size = 4 gradient_accumulation_steps = 4 max_seq_length = 4096
Full fine-tune (small models)	method = full learning_rate = 1e-05 epochs = 2 per_device_train_batch_size = 4 gradient_accumulation_steps = 4
CPU smoke test	method = lora lora_r = 8 lora_alpha = 16 max_steps = 20 epochs = 1 per_device_train_batch_size = 1 gradient_accumulation_steps = 1 max_seq_length = 256 precision = fp32 gradient_checkpointing = false

Preference tuning (DPO / ORPO / KTO)

Type train.preference · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Aligns a model with preferred answers: DPO, ORPO, SimPO or KTO with TRL. Usually run after SFT.

Inputs

Input	Label	Accepts	Link
train	Preference data	PreferenceDataset	required
model	Model	Model	required
eval	Validation data	PreferenceDataset	optional

Outputs

Output	Label	Type
model	Aligned model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
algorithm	Algorithm	select	dpo	dpo (DPO) orpo (ORPO (no reference model)) simpo (SimPO (reference-free)) kto (KTO (unpaired good/bad labels))	Always	—
beta	Beta	number	0.1	0 to 10, step 0.01	Always	Strength of the preference signal.
max_prompt_length	Max prompt length (tokens)	integer	1024	32 to 131072	Always	—
method	Method	select	qlora	qlora (QLoRA (4-bit base + LoRA)) lora (LoRA (16-bit base + LoRA)) full (Full fine-tune)	Always	How much of the model is trained.
epochs	Epochs	number	1	0.1 to 50, step 0.1	Always	—
max_steps	Max steps (-1 = use epochs)	integer	-1	-1 to 1000000, step 1	Always	Stop after this many optimizer steps. Handy for quick tests.
learning_rate	Learning rate	number	5e-06	1e-08 to 0.1, step 1e-06	Always	—
max_seq_length	Max sequence length (tokens)	integer	2048	64 to 262144, step 64	Always	Longer examples are truncated. Memory grows with this.

LoRA adapter

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lora_r	LoRA rank (r)	integer	16	1 to 512, step 1	method is lora or qlora	Adapter size. 8-32 covers most tasks. vLLM serves ranks up to its --max-lora-rank (default 16).
lora_alpha	LoRA alpha	integer	32	1 to 1024, step 1	method is lora or qlora	Scaling factor. Common choice: 2 x rank.
lora_dropout	LoRA dropout	number	0.0	0 to 0.5, step 0.01	method is lora or qlora	—
target_modules	Target layers	select	all-linear	all-linear (All linear layers) attention (Attention only (q, k, v, o)) custom (Custom list)	method is lora or qlora	—
custom_target_modules	Custom target modules	text	(empty)	—	target_modules is custom	Example: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
use_dora	Use DoRA	boolean	false	—	method is lora or qlora	Weight-decomposed LoRA. Sometimes better quality, slower.
use_rslora	Use rank-stabilized LoRA	boolean	false	—	method is lora or qlora	Scales by alpha/sqrt(r); helps at high ranks.

Batch and memory

Setting	Label	Kind	Default	Choices / range	Shown when	Help
per_device_train_batch_size	Batch size per GPU	integer	2	1 to 512, step 1	Always	—
gradient_accumulation_steps	Gradient accumulation	integer	8	1 to 512, step 1	Always	Effective batch = batch size x accumulation x GPUs.
gradient_checkpointing	Gradient checkpointing	boolean	true	—	Always	Saves a lot of memory for ~20-30% more time.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lr_scheduler_type	LR schedule	select	cosine	cosine linear constant constant_with_warmup	Always	—
warmup_ratio	Warmup ratio	number	0.05	0 to 0.5, step 0.01	Always	—
weight_decay	Weight decay	number	0.01	0 to 1, step 0.001	Always	—
max_grad_norm	Max gradient norm	number	1.0	0 to 100, step 0.1	Always	—
optim	Optimizer	select	auto	auto (Auto) adamw_torch_fused (AdamW (fused)) adamw_torch (AdamW) adamw_8bit (AdamW 8-bit) paged_adamw_8bit (Paged AdamW 8-bit) adafactor (Adafactor)	Always	—
precision	Precision	select	auto	auto (Auto) bf16 fp16 fp32	Always	—
seed	Seed	integer	42	0 to 2147483647, step 1	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
logging_steps	Log every N steps	integer	5	1 to 10000	Always	Not part of the cache key: changing it does not re-run a cached step.
eval_strategy	Evaluate during training	select	epoch	no steps epoch	Always	Only used when the eval input is connected.
eval_steps	Eval every N steps	integer	100	1 to 1000000	eval_strategy is steps	—
save_steps	Checkpoint every N steps	integer	200	1 to 1000000	Always	Checkpoints let a stopped or preempted job resume. Not part of the cache key: changing it does not re-run a cached step.
early_stopping_patience	Early stopping patience (0 = off)	integer	0	0 to 100	Always	Stop when eval loss fails to improve this many evaluations in a row.
report_to	Also log to	select	none	none (Tuneline only) wandb (Weights & Biases)	Always	Not part of the cache key: changing it does not re-run a cached step.

Speed-ups

Setting	Label	Kind	Default	Choices / range	Shown when	Help
use_unsloth	Use Unsloth	boolean	false	—	Always	Faster, lower-memory training on one NVIDIA GPU for supported models. Current Unsloth releases need an older TRL than Tuneline uses, so until Unsloth supports TRL 1 this falls back to standard TRL (logged in the run).
use_liger_kernel	Use Liger kernels	boolean	false	—	Always	Fused kernels that cut memory for supported architectures.

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the adapter to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-aligned	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

Choices for method:

• qlora — QLoRA (4-bit base + LoRA): Lowest memory. Good default for 7B+ models on one GPU.
• lora — LoRA (16-bit base + LoRA): Slightly better quality, about 2-3x the memory of QLoRA.
• full — Full fine-tune: Updates every weight. Needs ~16-18 bytes per parameter; small models only.

Choices for target_modules:

• all-linear — All linear layers: Recommended (attention + MLP).
• attention — Attention only (q, k, v, o)
• custom — Custom list

Choices for optim:

• auto — Auto: paged 8-bit AdamW for QLoRA, fused AdamW otherwise
• adamw_torch_fused — AdamW (fused)
• adamw_torch — AdamW
• adamw_8bit — AdamW 8-bit
• paged_adamw_8bit — Paged AdamW 8-bit
• adafactor — Adafactor

Choices for precision:

• auto — Auto: bf16 on Ampere or newer, fp16 on T4/V100, fp32 on CPU
• bf16
• fp16
• fp32

API keys

Key	What it is	Need	When	Why
wandb_api_key	Weights & Biases API key	required	When report_to is wandb	Log training to Weights & Biases
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the adapter

Presets

Choosing a preset in the inspector replaces these settings; the rest keep their values.

Preset	Values
DPO with LoRA	algorithm = dpo method = lora learning_rate = 5e-05 beta = 0.1 epochs = 1
DPO full (small models)	algorithm = dpo method = full learning_rate = 1e-06 beta = 0.1
ORPO with QLoRA	algorithm = orpo method = qlora learning_rate = 8e-06 beta = 0.1

Vision fine-tune (SFT)

Type train.vision_sft · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Fine-tunes a vision-language model (Qwen-VL, Gemma 3/4, Llama 3.2 Vision, SmolVLM and others) on image + text chats, for documents, receipts, IDs and screenshots.

Inputs

Input	Label	Accepts	Link
train	Training data	VisionDataset	required
model	Vision model	Model	required
eval	Validation data	VisionDataset	optional

Outputs

Output	Label	Type
model	Fine-tuned model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
method	Method	select	qlora	qlora (QLoRA (4-bit base + LoRA)) lora (LoRA (16-bit base + LoRA)) full (Full fine-tune)	Always	How much of the model is trained.
epochs	Epochs	number	1	0.1 to 50, step 0.1	Always	—
max_steps	Max steps (-1 = use epochs)	integer	-1	-1 to 1000000, step 1	Always	Stop after this many optimizer steps. Handy for quick tests.
learning_rate	Learning rate	number	0.0001	1e-08 to 0.1, step 1e-06	Always	—
max_seq_length	Max sequence length (tokens)	integer	2048	64 to 262144, step 64	Always	Longer examples are truncated. Memory grows with this.

LoRA adapter

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lora_r	LoRA rank (r)	integer	16	1 to 512, step 1	method is lora or qlora	Adapter size. 8-32 covers most tasks. vLLM serves ranks up to its --max-lora-rank (default 16).
lora_alpha	LoRA alpha	integer	32	1 to 1024, step 1	method is lora or qlora	Scaling factor. Common choice: 2 x rank.
lora_dropout	LoRA dropout	number	0.0	0 to 0.5, step 0.01	method is lora or qlora	—
target_modules	Target layers	select	all-linear	all-linear (All linear layers) attention (Attention only (q, k, v, o)) custom (Custom list)	method is lora or qlora	—
custom_target_modules	Custom target modules	text	(empty)	—	target_modules is custom	Example: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
use_dora	Use DoRA	boolean	false	—	method is lora or qlora	Weight-decomposed LoRA. Sometimes better quality, slower.
use_rslora	Use rank-stabilized LoRA	boolean	false	—	method is lora or qlora	Scales by alpha/sqrt(r); helps at high ranks.
finetune_vision_layers	Also train the vision encoder	boolean	false	—	Always	—

Batch and memory

Setting	Label	Kind	Default	Choices / range	Shown when	Help
per_device_train_batch_size	Batch size per GPU	integer	1	1 to 512, step 1	Always	—
gradient_accumulation_steps	Gradient accumulation	integer	8	1 to 512, step 1	Always	Effective batch = batch size x accumulation x GPUs.
gradient_checkpointing	Gradient checkpointing	boolean	true	—	Always	Saves a lot of memory for ~20-30% more time.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lr_scheduler_type	LR schedule	select	cosine	cosine linear constant constant_with_warmup	Always	—
warmup_ratio	Warmup ratio	number	0.05	0 to 0.5, step 0.01	Always	—
weight_decay	Weight decay	number	0.01	0 to 1, step 0.001	Always	—
max_grad_norm	Max gradient norm	number	1.0	0 to 100, step 0.1	Always	—
optim	Optimizer	select	auto	auto (Auto) adamw_torch_fused (AdamW (fused)) adamw_torch (AdamW) adamw_8bit (AdamW 8-bit) paged_adamw_8bit (Paged AdamW 8-bit) adafactor (Adafactor)	Always	—
precision	Precision	select	auto	auto (Auto) bf16 fp16 fp32	Always	—
seed	Seed	integer	42	0 to 2147483647, step 1	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
logging_steps	Log every N steps	integer	5	1 to 10000	Always	Not part of the cache key: changing it does not re-run a cached step.
eval_strategy	Evaluate during training	select	epoch	no steps epoch	Always	Only used when the eval input is connected.
eval_steps	Eval every N steps	integer	100	1 to 1000000	eval_strategy is steps	—
save_steps	Checkpoint every N steps	integer	200	1 to 1000000	Always	Checkpoints let a stopped or preempted job resume. Not part of the cache key: changing it does not re-run a cached step.
early_stopping_patience	Early stopping patience (0 = off)	integer	0	0 to 100	Always	Stop when eval loss fails to improve this many evaluations in a row.
report_to	Also log to	select	none	none (Tuneline only) wandb (Weights & Biases)	Always	Not part of the cache key: changing it does not re-run a cached step.

Speed-ups

Setting	Label	Kind	Default	Choices / range	Shown when	Help
use_unsloth	Use Unsloth	boolean	false	—	Always	Faster, lower-memory training on one NVIDIA GPU for supported models. Current Unsloth releases need an older TRL than Tuneline uses, so until Unsloth supports TRL 1 this falls back to standard TRL (logged in the run).

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the adapter to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-vision-adapter	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

Choices for method:

• qlora — QLoRA (4-bit base + LoRA): Lowest memory. Good default for 7B+ models on one GPU.
• lora — LoRA (16-bit base + LoRA): Slightly better quality, about 2-3x the memory of QLoRA.
• full — Full fine-tune: Updates every weight. Needs ~16-18 bytes per parameter; small models only.

Choices for target_modules:

• all-linear — All linear layers: Recommended (attention + MLP).
• attention — Attention only (q, k, v, o)
• custom — Custom list

Choices for optim:

• auto — Auto: paged 8-bit AdamW for QLoRA, fused AdamW otherwise
• adamw_torch_fused — AdamW (fused)
• adamw_torch — AdamW
• adamw_8bit — AdamW 8-bit
• paged_adamw_8bit — Paged AdamW 8-bit
• adafactor — Adafactor

Choices for precision:

• auto — Auto: bf16 on Ampere or newer, fp16 on T4/V100, fp32 on CPU
• bf16
• fp16
• fp32

API keys

Key	What it is	Need	When	Why
wandb_api_key	Weights & Biases API key	required	When report_to is wandb	Log training to Weights & Biases
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the adapter

Presets

Choosing a preset in the inspector replaces these settings; the rest keep their values.

Preset	Values
QLoRA vision quick	method = qlora lora_r = 16 lora_alpha = 32 learning_rate = 0.0001 epochs = 1 per_device_train_batch_size = 1 gradient_accumulation_steps = 8

Reinforcement tuning (GRPO)

Type train.grpo · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

GRPO with rule-based rewards (exact answer, contains answer, format regex, valid JSON, length). Good for tasks with checkable answers. Needs prompt/completion rows where the completion is the reference answer.

Inputs

Input	Label	Accepts	Link
train	Prompts with reference answers	ChatDataset	required
model	Model	Model	required

Outputs

Output	Label	Type
model	Tuned model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
rewards	Rewards	multiselect	["contains_answer", "format_regex"]	exact_answer (Exact answer) contains_answer (Contains the answer) format_regex (Matches a format regex) json_valid (Valid JSON) length_penalty (Penalize long answers)	Always	—
format_regex	Format regex	text	(?s).*<answer>.*</answer>.*	—	rewards includes format_regex	—
answer_regex	Extract answer with regex (group 1)	text	<answer>(.*?)</answer>	—	Always	—
num_generations	Generations per prompt	integer	8	2 to 64	Always	—
max_completion_length	Max completion tokens	integer	512	16 to 32768	Always	—
beta	KL beta	number	0.0	0 to 1, step 0.01	Always	—
method	Method	select	qlora	qlora (QLoRA (4-bit base + LoRA)) lora (LoRA (16-bit base + LoRA)) full (Full fine-tune)	Always	How much of the model is trained.
epochs	Epochs	number	1	0.1 to 50, step 0.1	Always	—
max_steps	Max steps (-1 = use epochs)	integer	-1	-1 to 1000000, step 1	Always	Stop after this many optimizer steps. Handy for quick tests.
learning_rate	Learning rate	number	1e-05	1e-08 to 0.1, step 1e-06	Always	—
max_seq_length	Max sequence length (tokens)	integer	1024	64 to 262144, step 64	Always	Longer examples are truncated. Memory grows with this.

LoRA adapter

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lora_r	LoRA rank (r)	integer	16	1 to 512, step 1	method is lora or qlora	Adapter size. 8-32 covers most tasks. vLLM serves ranks up to its --max-lora-rank (default 16).
lora_alpha	LoRA alpha	integer	32	1 to 1024, step 1	method is lora or qlora	Scaling factor. Common choice: 2 x rank.
lora_dropout	LoRA dropout	number	0.0	0 to 0.5, step 0.01	method is lora or qlora	—
target_modules	Target layers	select	all-linear	all-linear (All linear layers) attention (Attention only (q, k, v, o)) custom (Custom list)	method is lora or qlora	—
custom_target_modules	Custom target modules	text	(empty)	—	target_modules is custom	Example: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
use_dora	Use DoRA	boolean	false	—	method is lora or qlora	Weight-decomposed LoRA. Sometimes better quality, slower.
use_rslora	Use rank-stabilized LoRA	boolean	false	—	method is lora or qlora	Scales by alpha/sqrt(r); helps at high ranks.

Batch and memory

Setting	Label	Kind	Default	Choices / range	Shown when	Help
per_device_train_batch_size	Batch size per GPU	integer	8	1 to 512, step 1	Always	—
gradient_accumulation_steps	Gradient accumulation	integer	1	1 to 512, step 1	Always	Effective batch = batch size x accumulation x GPUs.
gradient_checkpointing	Gradient checkpointing	boolean	true	—	Always	Saves a lot of memory for ~20-30% more time.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
lr_scheduler_type	LR schedule	select	cosine	cosine linear constant constant_with_warmup	Always	—
warmup_ratio	Warmup ratio	number	0.05	0 to 0.5, step 0.01	Always	—
weight_decay	Weight decay	number	0.01	0 to 1, step 0.001	Always	—
max_grad_norm	Max gradient norm	number	1.0	0 to 100, step 0.1	Always	—
optim	Optimizer	select	auto	auto (Auto) adamw_torch_fused (AdamW (fused)) adamw_torch (AdamW) adamw_8bit (AdamW 8-bit) paged_adamw_8bit (Paged AdamW 8-bit) adafactor (Adafactor)	Always	—
precision	Precision	select	auto	auto (Auto) bf16 fp16 fp32	Always	—
seed	Seed	integer	42	0 to 2147483647, step 1	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
logging_steps	Log every N steps	integer	5	1 to 10000	Always	Not part of the cache key: changing it does not re-run a cached step.
eval_strategy	Evaluate during training	select	epoch	no steps epoch	Always	Only used when the eval input is connected.
eval_steps	Eval every N steps	integer	100	1 to 1000000	eval_strategy is steps	—
save_steps	Checkpoint every N steps	integer	200	1 to 1000000	Always	Checkpoints let a stopped or preempted job resume. Not part of the cache key: changing it does not re-run a cached step.
early_stopping_patience	Early stopping patience (0 = off)	integer	0	0 to 100	Always	Stop when eval loss fails to improve this many evaluations in a row.
report_to	Also log to	select	none	none (Tuneline only) wandb (Weights & Biases)	Always	Not part of the cache key: changing it does not re-run a cached step.

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the adapter to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-grpo	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

Choices for method:

• qlora — QLoRA (4-bit base + LoRA): Lowest memory. Good default for 7B+ models on one GPU.
• lora — LoRA (16-bit base + LoRA): Slightly better quality, about 2-3x the memory of QLoRA.
• full — Full fine-tune: Updates every weight. Needs ~16-18 bytes per parameter; small models only.

Choices for target_modules:

• all-linear — All linear layers: Recommended (attention + MLP).
• attention — Attention only (q, k, v, o)
• custom — Custom list

Choices for optim:

• auto — Auto: paged 8-bit AdamW for QLoRA, fused AdamW otherwise
• adamw_torch_fused — AdamW (fused)
• adamw_torch — AdamW
• adamw_8bit — AdamW 8-bit
• paged_adamw_8bit — Paged AdamW 8-bit
• adafactor — Adafactor

Choices for precision:

• auto — Auto: bf16 on Ampere or newer, fp16 on T4/V100, fp32 on CPU
• bf16
• fp16
• fp32

API keys

Key	What it is	Need	When	Why
wandb_api_key	Weights & Biases API key	required	When report_to is wandb	Log training to Weights & Biases
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the adapter

Train a segmenter (U-Net)

Type train.segmenter · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Trains a U-Net that labels every pixel with a class (for example document fields) and, optionally, a separate binary mask (for example tampered or defective areas).

Inputs

Input	Label	Accepts	Link
train	Segmentation arrays	ImageData	required

Outputs

Output	Label	Type
model	Segmenter	VisionNet
report	Validation report	EvalReport

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
encoder	Encoder	select	mobilenet_v2	mobilenet_v2 (MobileNetV2) resnet18 (ResNet-18) resnet34 (ResNet-34) efficientnet-b0 (EfficientNet-B0)	Always	—
classes	Number of classes (incl. background)	integer	8	2 to 256	Always	—
mask_head	Extra binary mask head	boolean	true	—	Always	Adds one more output trained on m.npy (for example tampered pixels).
epochs	Epochs	integer	14	1 to 500	Always	—
batch_size	Batch size	integer	6	1 to 1024	Always	—
learning_rate	Learning rate	number	0.0005	1e-06 to 0.1, step 1e-05	Always	—
max_steps_per_epoch	Max steps per epoch (0 = all)	integer	0	0 to 1000000	Always	Caps each epoch. Handy for quick tests.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
pretrained	Start from ImageNet weights	boolean	true	—	Always	Downloads the backbone's ImageNet weights once. Usually trains faster and better.
weight_decay	Weight decay	number	0.0001	0 to 1, step 1e-05	Always	—
warmup_steps	Warmup steps	integer	300	0 to 100000	Always	—
amp	Mixed precision (fp16) on the GPU	boolean	true	—	Always	—
workers	Data loader workers	integer	4	0 to 64	Always	Not part of the cache key: changing it does not re-run a cached step.
seed	Seed	integer	42	0 to 2147483647	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
checkpoint_every	Checkpoint every N steps	integer	400	10 to 1000000	Always	A stopped run resumes from the last checkpoint when it runs again with the same settings. Not part of the cache key: changing it does not re-run a cached step.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
background_weight	Background class weight	number	0.15	0 to 1, step 0.01	Always	—
mask_pos_weight	Mask positive weight	number	4.0	0.1 to 100, step 0.1	mask_head is true	—
shift_augment	Random shifts	boolean	true	—	Always	—

Choices for encoder:

• mobilenet_v2 — MobileNetV2: Small and fast on CPU after export (about 26 MB).
• resnet18 — ResNet-18
• resnet34 — ResNet-34
• efficientnet-b0 — EfficientNet-B0

API keys

None.

Train a text reader (CRNN)

Type train.text_reader · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Trains a text-line recogniser (convolutions + BiLSTM, CTC loss) on image crops with their text: handwriting, printed fields or MICR lines.

Inputs

Input	Label	Accepts	Link
train	Image shards with text	ImageData	required

Outputs

Output	Label	Type
model	Text reader	VisionNet
report	Validation report	EvalReport

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
charset	Characters	select	latin_micr	latin_micr (Letters, digits, punctuation, MICR symbols) latin (Letters, digits, punctuation) digits (Digits and . , - /)	Always	—
height	Input height (px)	integer	40	16 to 128	Always	—
width	Input width (px)	integer	1024	64 to 4096	Always	—
epochs	Epochs	integer	10	1 to 500	Always	—
batch_size	Batch size	integer	64	1 to 1024	Always	—
learning_rate	Learning rate	number	0.001	1e-06 to 0.1, step 1e-05	Always	—
max_steps_per_epoch	Max steps per epoch (0 = all)	integer	0	0 to 1000000	Always	Caps each epoch. Handy for quick tests.

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
pretrained	Start from ImageNet weights	boolean	true	—	Always	Downloads the backbone's ImageNet weights once. Usually trains faster and better.
weight_decay	Weight decay	number	0.0001	0 to 1, step 1e-05	Always	—
warmup_steps	Warmup steps	integer	300	0 to 100000	Always	—
amp	Mixed precision (fp16) on the GPU	boolean	true	—	Always	—
workers	Data loader workers	integer	4	0 to 64	Always	Not part of the cache key: changing it does not re-run a cached step.
seed	Seed	integer	42	0 to 2147483647	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
checkpoint_every	Checkpoint every N steps	integer	400	10 to 1000000	Always	A stopped run resumes from the last checkpoint when it runs again with the same settings. Not part of the cache key: changing it does not re-run a cached step.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
hidden	LSTM size	integer	192	32 to 1024	Always	—
skip_tags	Skip rows tagged (comma separated)	text	signature	—	Always	Rows whose tag is listed are not used (for example signatures, which are not read).

API keys

None.

Train an image embedder (similarity)

Type train.embedder · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Learns an embedding where images of the same group are close and others far apart (supervised contrastive or batch-hard triplet loss). Use it to verify writers, signatures, products or documents.

Inputs

Input	Label	Accepts	Link
train	Image shards with groups	ImageData	required

Outputs

Output	Label	Type
model	Embedder	VisionNet
report	Validation report	EvalReport

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
backbone	Backbone	select	mobilenetv3_large_100	mobilenetv3_large_100 (MobileNetV3-Large) resnet18 (ResNet-18) efficientnet_b0 (EfficientNet-B0)	Always	—
dim	Embedding size	integer	128	8 to 2048	Always	—
height	Input height (px)	integer	64	16 to 1024	Always	—
width	Input width (px)	integer	256	16 to 2048	Always	—
fit	Resize	select	letterbox	letterbox (Keep aspect, pad) stretch (Stretch to size)	Always	—
groups_per_batch	Groups per batch	integer	40	2 to 1024	Always	—
items_per_group	Images per group	integer	4	2 to 64	Always	—
loss	Loss	select	supcon	supcon (Supervised contrastive) batch_hard (Batch-hard triplet)	Always	—
margin	Triplet margin (batch-hard)	number	0.25	0.01 to 2, step 0.01	Always	—
steps_per_epoch	Steps per epoch	integer	500	1 to 1000000	Always	—
epochs	Epochs	integer	10	1 to 500	Always	—
learning_rate	Learning rate	number	0.0005	1e-06 to 0.1, step 1e-05	Always	—

Optimizer

Setting	Label	Kind	Default	Choices / range	Shown when	Help
pretrained	Start from ImageNet weights	boolean	true	—	Always	Downloads the backbone's ImageNet weights once. Usually trains faster and better.
weight_decay	Weight decay	number	0.0001	0 to 1, step 1e-05	Always	—
warmup_steps	Warmup steps	integer	300	0 to 100000	Always	—
amp	Mixed precision (fp16) on the GPU	boolean	true	—	Always	—
workers	Data loader workers	integer	4	0 to 64	Always	Not part of the cache key: changing it does not re-run a cached step.
seed	Seed	integer	42	0 to 2147483647	Always	—

Evaluation and checkpoints

Setting	Label	Kind	Default	Choices / range	Shown when	Help
checkpoint_every	Checkpoint every N steps	integer	400	10 to 1000000	Always	A stopped run resumes from the last checkpoint when it runs again with the same settings. Not part of the cache key: changing it does not re-run a cached step.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
temperature	Temperature (supervised contrastive)	number	0.1	0.01 to 1, step 0.01	Always	—
include_tags	Only rows tagged (comma separated, blank = all)	text	(empty)	—	Always	—

Choices for backbone:

• mobilenetv3_large_100 — MobileNetV3-Large: Small and fast (about 17 MB).
• resnet18 — ResNet-18
• efficientnet_b0 — EfficientNet-B0

Choices for loss:

• supcon — Supervised contrastive: Stable from scratch: pulls each group together against every other image in the batch.
• batch_hard — Batch-hard triplet: Hardest positive and negative per image. Can collapse (every image maps to the same point) on hard data.

API keys

None.

Evaluate

Evaluate on test set

Type eval.golden · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Generates answers for the test set and scores them: exact match, contains, label accuracy and F1, valid JSON, regex match, ROUGE-L and an optional LLM judge. Can score the base model too for comparison.

Inputs

Input	Label	Accepts	Link
test	Test data	ChatDataset, VisionDataset	required
model	Model	Model	required

Outputs

Output	Label	Type
report	Report	EvalReport

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
metrics	Metrics	multiselect	["exact_match", "rouge_l"]	exact_match (Exact match) contains (Contains expected) label_accuracy (Label accuracy and F1 (label regex)) json_valid (Valid JSON) regex_match (Matches regex) rouge_l (ROUGE-L) llm_judge (LLM judge (1-10))	Always	—
label_regex	Label regex (group 1 = label)	text	(?i)decision:\s*([A-Za-z_\-]+)	—	metrics includes label_accuracy	—
match_regex	Regex the answer must match	text	(empty)	—	metrics includes regex_match	—
max_samples	Max examples (0 = all)	integer	200	0 to 1000000	Always	—
max_new_tokens	Max new tokens	integer	256	1 to 32768	Always	—
temperature	Temperature	number	0.0	0 to 2, step 0.05	Always	—
batch_size	Generation batch size	integer	8	1 to 512	Always	Not part of the cache key: changing it does not re-run a cached step.
compare_base	Also score the base model	boolean	true	—	Always	Runs the same test without the adapter so you can see the gain.
judge_provider	Judge API	select	hf_router	hf_router (Hugging Face Inference Providers) openai_compatible (OpenAI-compatible API)	metrics includes llm_judge	—
judge_model	Judge model	text	Qwen/Qwen3-235B-A22B-Instruct-2507	—	metrics includes llm_judge	—
judge_rubric	Judge rubric	textarea	Score the ANSWER from 1 to 10 for correctness and helpfulness compared with the REFERENCE. Reply with only the number.	—	metrics includes llm_judge	—

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
backend	Generation backend	select	transformers	transformers (Transformers) vllm (vLLM (faster; NVIDIA GPU with PyTorch 2.10+, installed on first use; otherwise Transformers))	Always	Not part of the cache key: changing it does not re-run a cached step.

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	When judge_provider is hf_router and metrics includes llm_judge	LLM judge
openai_compat_key	API key for an OpenAI-compatible endpoint (judge / synthetic data)	required	When judge_provider is openai_compatible and metrics includes llm_judge	LLM judge
openai_compat_base_url	Base URL for that OpenAI-compatible endpoint	required	When judge_provider is openai_compatible and metrics includes llm_judge	LLM judge

Benchmarks (lm-eval)

Type eval.benchmarks · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Runs standard benchmarks with EleutherAI's lm-evaluation-harness to check general ability did not regress (IFEval, GSM8K, MMLU-Pro, ARC, HellaSwag and more).

Inputs

Input	Label	Accepts	Link
model	Model	Model	required

Outputs

Output	Label	Type
report	Report	EvalReport

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
tasks	Benchmarks	multiselect	["ifeval"]	ifeval (IFEval (instruction following)) gsm8k (GSM8K (math)) mmlu_pro (MMLU-Pro (knowledge)) arc_challenge (ARC-Challenge) hellaswag (HellaSwag) truthfulqa_mc2 (TruthfulQA) winogrande (WinoGrande)	Always	—
limit	Examples per benchmark (0 = all)	integer	100	0 to 100000	Always	—
num_fewshot	Few-shot examples (-1 = benchmark default)	integer	-1	-1 to 25	Always	—
compare_base	Also score the base model	boolean	false	—	Always	—

API keys

None.

Prompt-injection test

Type eval.redteam · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Hides injection strings ("ignore previous instructions...") inside test inputs and checks whether the model's answers change or leak a canary.

Inputs

Input	Label	Accepts	Link
test	Test data	ChatDataset	required
model	Model	Model	required

Outputs

Output	Label	Type
report	Report	EvalReport

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
attacks_per_example	Attacks per example	integer	2	1 to 10	Always	—
max_samples	Max examples	integer	50	1 to 10000	Always	—
max_new_tokens	Max new tokens	integer	256	1 to 8192	Always	—

API keys

None.

Quality gate

Type eval.gate · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · always runs (never cached)

Passes the model on only if every rule holds, for example 'label_accuracy >= 0.95'. A failed gate stops everything after it.

Inputs

Input	Label	Accepts	Link
report	Report	EvalReport	required
model	Model	Model, VisionNet	required
extra_report	Second report	EvalReport	optional

Outputs

Output	Label	Type
model	Approved model	Same type as the input model (one of Model, VisionNet)

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
rules	Rules (one per line)	rules	exact_match >= 0.8	—	Always	metric op value, where op is >=, >, <=, < or ==. Metric names come from the reports; prefix with base. or delta. to use the base-model score or the gain.
on_fail	When a rule fails	select	fail_run	fail_run (Fail the run) skip_downstream (Skip later steps, keep the run successful)	Always	—

API keys

None.

Export

Merge adapter

Type export.merge · version 1.0 · runtime: gpu (needs a GPU for real models) · results are cached

Merges the LoRA adapter into the base weights and saves a standalone model.

Inputs

Input	Label	Accepts	Link
model	Model with adapter	Model	required

Outputs

Output	Label	Type
model	Merged model	Model

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
dtype	Saved precision	select	bf16	bf16 fp16 fp32	Always	—

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the merged model to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-merged	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the merged model

Export GGUF

Type export.gguf · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Converts the model to GGUF with llama.cpp and quantizes it (Q4_K_M, Q5_K_M, Q8_0 ...) for Ollama, LM Studio and llama.cpp.

Inputs

Input	Label	Accepts	Link
model	Model	Model	required

Outputs

Output	Label	Type
gguf	GGUF files	GGUF

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
quantizations	Quantizations	multiselect	["q4_k_m", "q8_0"]	f16 bf16 q8_0 q6_k q5_k_m q4_k_m q3_k_m q2_k	Always	—

Hugging Face Hub

Setting	Label	Kind	Default	Choices / range	Shown when	Help
push_to_hub	Push the GGUF files to Hugging Face	boolean	true	—	Always	Needed to reuse the result on another machine and to deploy it. Not part of the cache key: changing it does not re-run a cached step.
hub_repo	Repository name	text	{pipeline}-gguf	—	push_to_hub is true	Created under your Hugging Face account. {pipeline} and {run} are filled in. Not part of the cache key: changing it does not re-run a cached step.
hub_private	Private repository	boolean	true	—	push_to_hub is true	Not part of the cache key: changing it does not re-run a cached step.

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	When push_to_hub is true	Push the GGUF files

Publish and register

Type export.hub · version 1.0 · runtime: network (needs internet access only) · always runs (never cached)

Pushes the adapter or merged model to your Hugging Face account with a generated model card, and records it as a new version in your model registry.

Inputs

Input	Label	Accepts	Link
model	Model	Model	required
report	Eval report	EvalReport	optional

Outputs

Output	Label	Type
model	Published model	Model

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
hub_repo	Repository name (required)	text	{pipeline}	—	Always	—
hub_private	Private repository	boolean	true	—	Always	—
register	Add to model registry	boolean	true	—	Always	—
registry_name	Registry name	text	{pipeline}	—	register is true	—
card_notes	Notes for the model card	textarea	(empty)	—	Always	—

API keys

Key	What it is	Need	When	Why
hf_token	Hugging Face access token (write)	required	Always	Push to your Hugging Face account

Export to ONNX

Type export.onnx · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · results are cached

Exports a trained vision model to ONNX (runs on CPU with ONNX Runtime, no PyTorch), checks that ONNX and PyTorch agree and copies the file to a folder.

Inputs

Input	Label	Accepts	Link
model	Vision model	VisionNet	required

Outputs

Output	Label	Type
model	Exported model	VisionNet

Settings

Basics

Setting	Label	Kind	Default	Choices / range	Shown when	Help
out_dir	Copy to folder (optional)	text	(empty)	—	Always	Folder on the machine that runs the job; the file is always kept with the run too. Not part of the cache key: changing it does not re-run a cached step.
file_name	File name	text	{node}.onnx	—	Always	{node} is the trainer's step id. Not part of the cache key: changing it does not re-run a cached step.

Advanced

Setting	Label	Kind	Default	Choices / range	Shown when	Help
opset	ONNX opset	integer	17	11 to 22	Always	—
verify	Check ONNX against PyTorch	boolean	true	—	Always	—

API keys

None.

Deploy

Deploy endpoint

Type deploy.endpoint · version 1.0 · runtime: network (needs internet access only) · always runs (never cached)

Starts an OpenAI-compatible vLLM endpoint for the model on RunPod Serverless, Hugging Face Inference Endpoints or Modal. The model must be on the Hub (push or merge it first).

Inputs

Input	Label	Accepts	Link
model	Model	Model	required

Outputs

Output	Label	Type
endpoint	Endpoint	Endpoint

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
provider	Where to serve	select	runpod_serverless	runpod_serverless (RunPod Serverless (scales to zero)) hf_endpoint (Hugging Face Inference Endpoints (scales to zero)) modal (Modal (one vLLM container, stops after a time limit))	Always	—
name	Endpoint name	text	{pipeline}	—	Always	—
gpu	GPU	select	24GB	16GB (16 GB (T4 class)) 24GB (24 GB (L4 / A10 / 4090)) 48GB (48 GB (L40S / A6000)) 80GB (80 GB (A100 / H100))	Always	—
max_model_len	Max context length	integer	4096	256 to 262144	Always	—
min_workers	Always-on workers	integer	0	0 to 10	provider is runpod_serverless or hf_endpoint	—
max_workers	Max workers	integer	1	1 to 20	provider is runpod_serverless or hf_endpoint	—
idle_timeout_s	Scale down after idle (seconds)	integer	300	5 to 86400	provider is runpod_serverless or hf_endpoint	—
max_hours	Stop after (hours)	number	2	0.25 to 24, step 0.25	provider is modal	—

API keys

Key	What it is	Need	When	Why
runpod_api_key	RunPod API key	required	When provider is runpod_serverless	Create the RunPod endpoint
hf_token	Hugging Face access token (write)	required	Always	Let the server download your model
modal_token_id	Modal token ID	required	When provider is modal	Create the Modal app
modal_token_secret	Modal token secret	required	When provider is modal	Create the Modal app

Ollama package

Type deploy.ollama · version 1.0 · runtime: cpu (runs on a CPU; no GPU needed) · always runs (never cached)

Writes an Ollama Modelfile for the GGUF export. On your own machine with Ollama installed it also runs 'ollama create' so the model is ready to chat.

Inputs

Input	Label	Accepts	Link
gguf	GGUF files	GGUF	required

Outputs

Output	Label	Type
endpoint	Local endpoint	Endpoint

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
model_name	Ollama model name	text	{pipeline}	—	Always	—
quantization	Which file	select	q4_k_m	q4_k_m q5_k_m q8_0 f16 bf16	Always	—
system_prompt	System prompt	textarea	(empty)	—	Always	—
temperature	Temperature	number	0.2	0 to 2, step 0.05	Always	—

API keys

None.

Endpoint smoke test

Type deploy.smoke_test · version 1.0 · runtime: network (needs internet access only) · always runs (never cached)

Sends test prompts to the endpoint and records latency and answers.

Inputs

Input	Label	Accepts	Link
endpoint	Endpoint	Endpoint	required

Outputs

Output	Label	Type
report	Report	EvalReport

Settings

Setting	Label	Kind	Default	Choices / range	Shown when	Help
prompts	Test prompts (one per line)	textarea	Hello! What can you do?	—	Always	—
expect_contains	Answers must contain (optional)	text	(empty)	—	Always	—
max_latency_ms	Max latency (ms)	integer	30000	100 to 600000	Always	—
wait_for_ready_s	Wait for cold start (seconds)	integer	600	0 to 3600	Always	—

API keys

Key	What it is	Need	When	Why
runpod_api_key	RunPod API key	optional	Always	Call RunPod endpoints
hf_token	Hugging Face access token (write)	optional	Always	Call Hugging Face endpoints

API key names

The key names used above, as labelled in Settings → API keys:

Key	What it is
hf_token	Hugging Face access token (write)
modal_token_id	Modal token ID
modal_token_secret	Modal token secret
runpod_api_key	RunPod API key
kaggle_api_token	Kaggle API token
kaggle_username	Kaggle username (legacy key)
kaggle_key	Kaggle key (legacy key)
openai_compat_key	API key for an OpenAI-compatible endpoint (judge / synthetic data)
openai_compat_base_url	Base URL for that OpenAI-compatible endpoint
wandb_api_key	Weights & Biases API key