Templates

The 9 starting points offered by Pipelines → New pipeline: what each one does, its steps in run order with the settings the template changes, and the API keys it needs.

Note

Generated from the block registry and the templates 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.

A template is only a starting graph: once created, the pipeline is yours to edit. Settings not listed under Key settings keep the block's default (see the Block reference). Values with {app_url} (the app's address) and {pipeline} (your pipeline's slug) are filled in when the run starts. Long values are shortened here; the pipeline's YAML tab shows them in full.

Where you run adds its own keys: Modal needs its token id and secret, RunPod its API key, Hugging Face Jobs your Hugging Face token and Kaggle a Kaggle API token (or the legacy username and key); Notebook and My computer need none (see Hardware and costs).

Overview

Template	Id	Tags	Steps	Required API keys
Sanctions alert triage	sanctions-triage	text, classification, banking	10	hf_token, runpod_api_key
Chat assistant (SFT → GGUF → Ollama)	chat-sft	text, chat, local	9	hf_token
Preference tuning (DPO)	dpo-alignment	text, alignment	6	hf_token
Document and image understanding (vision)	vision-documents	vision, documents	7	hf_token
Reasoning with rewards (GRPO)	grpo-reasoning	text, reasoning, rl	6	hf_token
Compare base models	compare-models	evaluation	8	none
CPU smoke test (no GPU needed)	cpu-smoke-test	test	6	none
Document vision models (segmenter, text reader, embedder)	document-vision	vision, onnx, documents	11	none
Blank canvas	blank	—	0	none

Sanctions alert triage

Id sanctions-triage · tags: text, classification, banking

Fine-tunes a small model to decide ESCALATE or CLOSE on screening alerts, gates it on label accuracy and publishes it. Ships with a synthetic sample of 400 alerts.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = url url = {app_url}/samples/sanctions-alerts.jsonl	—
2	clean	Clean and dedupe	pii_redact = true	data ← load.data
3	chat	Format as chat	input_format = columns prompt_template = Alert: {alert_text}\nCustomer: {customer_name} (born {customer_dob}, {customer_country})… response_template = Decision: {decision}\nRationale: {rationale} system_prompt = You are an L1 sanctions screening analyst. Decide ESCALATE or CLOSE for the alert and giv… keep_columns = decision	data ← clean.data
4	base	Base model	repo_id = Qwen/Qwen3-4B-Instruct-2507	—
5	split	Train / validation / test split	test_size = 0.15 validation_size = 0.05 stratify_column = decision	data ← chat.data
6	sft	Fine-tune (SFT)	method = qlora epochs = 2 max_seq_length = 1024	train ← split.train eval ← split.validation model ← base.model
7	eval	Evaluate on test set	metrics = ["label_accuracy", "exact_match"] label_regex = (?i)decision:\s*([A-Za-z_\-]+) max_new_tokens = 128	test ← split.test model ← sft.model
8	gate	Quality gate	rules = label_accuracy >= 0.9\nlabel_recall_escalate >= 0.98	report ← eval.report model ← sft.model
9	publish	Publish and register	hub_repo = {pipeline} register = true	model ← gate.model report ← eval.report
10	deploy	Deploy endpoint	provider = runpod_serverless gpu = 24GB	model ← publish.model

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	required	sft: Push the adapter publish: Push to your Hugging Face account deploy: Let the server download your model
runpod_api_key	RunPod API key	required	deploy: Create the RunPod endpoint

Chat assistant (SFT → GGUF → Ollama)

Id chat-sft · tags: text, chat, local

Fine-tunes a small chat model on conversations from the Hub, compares it with the base model, merges it and exports GGUF files you can run in Ollama.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = hf hf_repo = HuggingFaceTB/smoltalk hf_config = everyday-conversations hf_split = train max_rows = 2000	—
2	chat	Format as chat	input_format = messages	data ← load.data
3	base	Base model	repo_id = Qwen/Qwen3-1.7B	—
4	split	Train / validation / test split	test_size = 0.05 validation_size = 0.02	data ← chat.data
5	sft	Fine-tune (SFT)	method = qlora epochs = 1	train ← split.train eval ← split.validation model ← base.model
6	eval	Evaluate on test set	metrics = ["rouge_l"] max_samples = 100 compare_base = true	test ← split.test model ← sft.model
7	merge	Merge adapter	defaults	model ← sft.model
8	gguf	Export GGUF	quantizations = ["q4_k_m", "q8_0"]	model ← merge.model
9	ollama	Ollama package	quantization = q4_k_m	gguf ← gguf.gguf

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	required	sft: Push the adapter merge: Push the merged model gguf: Push the GGUF files

Preference tuning (DPO)

Id dpo-alignment · tags: text, alignment

Aligns a model on chosen/rejected answer pairs with DPO, then publishes the adapter.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = hf hf_repo = trl-lib/ultrafeedback_binarized hf_split = train max_rows = 3000	—
2	pref	Format as preference pairs	input_format = paired	data ← load.data
3	base	Base model	repo_id = Qwen/Qwen3-1.7B	—
4	split	Train / validation / test split	test_size = 0.05 validation_size = 0.0	data ← pref.data
5	dpo	Preference tuning (DPO / ORPO / KTO)	algorithm = dpo method = lora learning_rate = 5e-05	train ← split.train eval ← split.test model ← base.model
6	publish	Publish and register	hub_repo = {pipeline} register = true	model ← dpo.model

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	required	dpo: Push the adapter publish: Push to your Hugging Face account

Document and image understanding (vision)

Id vision-documents · tags: vision, documents

Fine-tunes a vision-language model on image + question + answer rows, for example KYC documents, invoices or screenshots. Upload a ZIP with images and a JSONL file, or use a Hub dataset.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = upload	—
2	vision	Format as vision chat	input_format = columns image_column = image prompt_template = {question} response_template = {answer}	data ← load.data
3	base	Base model	repo_id = Qwen/Qwen3-VL-2B-Instruct modality = vision	—
4	split	Train / validation / test split	test_size = 0.1 validation_size = 0.0	data ← vision.data
5	sft	Vision fine-tune (SFT)	method = qlora	train ← split.train model ← base.model
6	eval	Evaluate on test set	metrics = ["exact_match", "contains"] max_samples = 100	test ← split.test model ← sft.model
7	publish	Publish and register	hub_repo = {pipeline} register = true	model ← sft.model report ← eval.report

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	required	sft: Push the adapter publish: Push to your Hugging Face account

Reasoning with rewards (GRPO)

Id grpo-reasoning · tags: text, reasoning, rl

Trains a small model on GSM8K math problems with rule-based rewards for the right answer and the answer format.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = hf hf_repo = openai/gsm8k hf_config = main hf_split = train max_rows = 2000	—
2	chat	Format as chat	input_format = columns prompt_template = {question}\n\nThink step by step, then give the final number inside <answer></answer>. response_template = {answer} output_style = prompt_completion	data ← load.data
3	base	Base model	repo_id = Qwen/Qwen3-1.7B	—
4	split	Train / validation / test split	test_size = 0.05 validation_size = 0.0	data ← chat.data
5	grpo	Reinforcement tuning (GRPO)	rewards = ["contains_answer", "format_regex"] max_steps = 300	train ← split.train model ← base.model
6	eval	Evaluate on test set	metrics = ["contains"] max_samples = 200 max_new_tokens = 512	test ← split.test model ← grpo.model

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	required	grpo: Push the adapter

Compare base models

Id compare-models · tags: evaluation

Scores two open models on your own test set and on standard benchmarks before you pick one to fine-tune.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = url url = {app_url}/samples/sanctions-alerts.jsonl	—
2	model_a	Base model	repo_id = Qwen/Qwen3-4B-Instruct-2507	—
3	model_b	Base model	repo_id = google/gemma-4-E4B-it	—
4	chat	Format as chat	input_format = columns prompt_template = Alert: {alert_text}\nCustomer: {customer_name} (born {customer_dob}, {customer_country})… response_template = Decision: {decision}\nRationale: {rationale} system_prompt = You are an L1 sanctions screening analyst. Decide ESCALATE or CLOSE for the alert and giv…	data ← load.data
5	eval_a	Evaluate on test set	metrics = ["label_accuracy"] compare_base = false	test ← chat.data model ← model_a.model
6	eval_b	Evaluate on test set	metrics = ["label_accuracy"] compare_base = false	test ← chat.data model ← model_b.model
7	bench_a	Benchmarks (lm-eval)	tasks = ["ifeval"] limit = 100	model ← model_a.model
8	bench_b	Benchmarks (lm-eval)	tasks = ["ifeval"] limit = 100	model ← model_b.model

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	optional	model_a: Gated models such as Llama and Gemma 2/3 model_b: Gated models such as Llama and Gemma 2/3

CPU smoke test (no GPU needed)

Id cpu-smoke-test · tags: test

Trains a 135M-parameter model for 20 steps on 60 sample chats. Runs on any computer in a few minutes, to check your setup end to end.

Steps

#	Step	Block	Key settings	Linked from
1	load	Load dataset	source = url url = {app_url}/samples/tiny-chat.jsonl	—
2	chat	Format as chat	input_format = messages	data ← load.data
3	base	Base model	repo_id = HuggingFaceTB/SmolLM2-135M-Instruct	—
4	split	Train / validation / test split	test_size = 0.1 validation_size = 0.0	data ← chat.data
5	sft	Fine-tune (SFT)	method = lora lora_r = 8 lora_alpha = 16 max_steps = 20 per_device_train_batch_size = 1 gradient_accumulation_steps = 1 max_seq_length = 256 precision = fp32 gradient_checkpointing = false push_to_hub = false	train ← split.train model ← base.model
6	eval	Evaluate on test set	metrics = ["rouge_l"] max_samples = 6 max_new_tokens = 48	test ← split.test model ← sft.model

API keys

Key	What it is	Need	Used by
hf_token	Hugging Face access token (write)	optional	base: Gated models such as Llama and Gemma 2/3

Document vision models (segmenter, text reader, embedder)

Id document-vision · tags: vision, onnx, documents

Trains three classic vision models from labelled images on your own GPU: a U-Net that finds fields and marks altered areas, a text reader for handwriting and printed lines, and an embedder that tells whether two images come from the same source. Each is gated on its validation metrics and exported to ONNX for CPU serving. Point the two image-data blocks at your folders first.

Steps

#	Step	Block	Key settings	Linked from
1	pages	Image data (folder)	path = /data/segmentation format = seg_cache val_split = test	—
2	crops	Image data (folder)	path = /data/crops format = image_shards val_shards = 2	—
3	segmenter	Train a segmenter (U-Net)	classes = 8 mask_head = true	train ← pages.data
4	reader	Train a text reader (CRNN)	charset = latin_micr	train ← crops.data
5	embedder	Train an image embedder (similarity)	defaults	train ← crops.data
6	gate_seg	Quality gate	rules = miou >= 0.6	report ← segmenter.report model ← segmenter.model
7	gate_reader	Quality gate	rules = cer <= 0.1	report ← reader.report model ← reader.model
8	gate_emb	Quality gate	rules = auc_random >= 0.9	report ← embedder.report model ← embedder.model
9	onnx_seg	Export to ONNX	file_name = segmenter	model ← gate_seg.model
10	onnx_reader	Export to ONNX	file_name = text_reader	model ← gate_reader.model
11	onnx_emb	Export to ONNX	file_name = embedder	model ← gate_emb.model

API keys

None for the pipeline itself.

Blank canvas

Id blank

Start from an empty pipeline.

An empty canvas: add blocks from the palette on the left.