OpenReef product guide

Train open models on a living reef of GPUs.

OpenReef is an AI lab built on the OpenGPU Network:

a decentralized reef of NVIDIA and AMD GPUs where you fine-tune open models on

your own data. No cluster, no YAML, no terminal — upload a dataset, pick a

model, confirm a price, download your adapter.

Explore. Tune. Create.

What is OpenReef?

Fine-tuning today usually means renting a cloud GPU, writing training configs,

watching logs and guessing the bill. OpenReef replaces all of that with a

four-step product flow:

Step	You do	OpenReef does
1	Upload a dataset	Inspects, maps columns and builds a clean, canonical training file — before you spend a credit
2	Pick a base model	Prices the job up front, in plain USD
3	Confirm	Routes the job to an eligible GPU on the OpenGPU Network
4	Wait a bit	Trains, validates the result and hands you a secure adapter download

You work with datasets, models and USD credits. The GPU reef underneath —

discovery, training and payment to providers — is handled for you.

Why OpenReef?

Traditional fine-tuning	OpenReef
Rent a GPU by the hour, configured or not	Pay only for the job, priced before you confirm
Training configs and framework flags by hand	Three curated modes with sane, hardware-aware defaults
Discover bad data after a failed (paid) run	Dataset Lab validates and cleans data before any job
Opaque cloud bills	Phase-aware refunds: a job that fails early returns the unused portion
One cloud provider	A decentralized network of independent GPU providers

Built for indie builders, researchers, students and small teams who want

open-source models on their own data — without the enterprise infra bill.

How it works

graph LR
    A["📦 Dataset Lab<br/>JSONL · CSV · TXT"] --> B["🧬 Model + mode<br/>Fast · Balanced · Custom"]
    B --> C["💠 Confirm price<br/>shown up front"]
    C --> D["⚡ SmartRoute<br/>eligible NVIDIA / AMD GPU"]
    D --> E["🏁 Adapter<br/>validated · presigned download"]

1. The Dataset Lab checks your data first

Training fails most often because of data, not models. The Dataset Lab

inspects every upload before a job can spend credits:

• Formats: JSONL (messages, ShareGPT, Alpaca, prompt/completion), CSV

and plain text are detected automatically.

• Column mapping: known aliases (instruction, pregunta, output,

respuesta, …) are mapped for you; unknown columns ask for a one-click

mapping.

• Quality report: row issues, warnings, token and turn distributions, and

a before/after preview of what will actually train. The full report can be

exported as JSON.

• Cleaning: empty and unusable rows are dropped; exact duplicates are

removed by default (you can keep them with one toggle).

• Canonical export: everything is normalized into one clean messages

JSONL format that training understands.

The whole pipeline is deterministic — no LLM ever reads your data here.

Every dataset is fingerprinted with SHA-256 so you always know exactly what

trained.

2. Pick a model

Browse the built-in catalog or import any open model from Hugging Face

(gated models work with your personal Hugging Face token). OpenReef knows each

model's chat template and trains the assistant side of each conversation.

3. Choose a mode and confirm the price

Mode	Promise
Fast	Low-cost iteration and smoke training
Balanced	The recommended baseline
Custom	Advanced controls (epochs, learning rate, LoRA rank and more) for technical users

The price is estimated before you confirm and depends on model size and

your configuration. Hardware and training engines stay invisible — OpenReef

picks an eligible GPU for the mode you chose.

4. Training and delivery

SmartRoute selects a provider by compatibility, VRAM, environment and

reputation. The job trains on that GPU with LoRA (or QLoRA), emits live

progress, and is validated before delivery: the final adapter is loaded

again and must produce sane output on held-out prompts. Only then can you

download the adapter zip from a secure, expiring link.

If a provider dies mid-job, OpenReef detects it by heartbeat and either

re-routes the job or refunds it — you never pay for silence.

Dataset Lab, in detail

Concern	Behavior
Upload	JSONL, CSV or TXT, up to 100 MB per file, 500 MB total per account
Analysis	Runs asynchronously in seconds-to-minutes; the page updates live
Verdicts	valid (trainable), needs_mapping (confirm columns), invalid (hard errors or fewer than 3 usable rows)
Provenance	SHA-256 of the original and the canonical file; jobs snapshot the fingerprint at creation
Immutability	Once a job uses a dataset, its mapping and canonical file are locked — upload a new version to change anything
Quality report	Row issues, warnings, token/turn distributions, raw before/after preview, exportable as JSON
Duplicates	Exact duplicates dropped by default; keep them with drop duplicates off
Retention	Failed datasets are warned at 7 days since last activity and cleaned after a grace period; unused datasets are warned at 60 days. Rename, retry analysis or save a new mapping cancels deletion and restarts the clock

New training jobs require a valid dataset — that is what stops failed runs

before they cost money.

Pricing and credits

• Prices are quoted before confirmation and charged in USD credits.
• Pricing follows the live OGPU/USD oracle; the credit ledger is append-only.
• Phase-aware refunds: a job that fails early returns the unused portion,

automatically.

• Pay with $OGPU on OpenGPU mainnet and receive +20% bonus credits.
• No subscriptions, no idle GPU tax: you pay per job, when you decide.

For GPU providers

Bring your NVIDIA or AMD card to the reef and earn OGPU on idle compute:

• Providers receive 50% of each job, settled on-chain in OGPU.
• Worker images are public and pinned by digest:

ghcr.io/asphyksia/finetune-worker (CUDA and ROCm).

• Claims are signed; the training container never holds your private key.
• You can pause your node at any time; in-flight jobs are protected by the

liveness and refund machinery.

Start here: provider onboarding.

Your data, protected

• Deterministic pipeline — dataset analysis uses no LLM; your rows are

inspected and reformatted, never generated or shared.

• Ephemeral access — dataset and adapter downloads are short-lived,

user-bound signed URLs. Nothing is publicly browsable.

• Provenance — SHA-256 fingerprints tie every job to the exact dataset it

trained on.

• Retention — unused or failed datasets are warned in-app and removed on a

published schedule; anything referenced by a job is kept immutable.

• Full terms: Terms of Service ·

Privacy Policy.

Limits (today)

Limit	Value
Active jobs per user	1
Dataset size	100 MB per file · 500 MB total
Dataset rows	100,000
Single source row	16 MB
Minimum usable rows	3 (warn below 50)
Sequence length	up to 4096 (hardware-dependent clamp)
Hugging Face import	models up to 14B parameters
Export	Adapter ZIP (merged weights and GGUF are on the roadmap)

FAQ

What do I get at the end?

A LoRA adapter ZIP with training metrics, validated by a serve smoke test.

Drop it into your inference stack or merge it into the base model.

Which models can I train?

The curated catalog plus any open Hugging Face model up to 14B parameters.

Gated families (e.g. Llama) work with your personal HF token.

How long does a job take?

From a few minutes (Fast on small models) to a few hours (Balanced/Custom on

larger models). The estimate appears with the price before you confirm.

What happens if my job fails?

You are refunded the unused portion automatically. GPU faults are never

blind-retried against your balance.

Is my dataset public?

No. Datasets are private to your account, served through expiring signed

URLs, and covered by the privacy policy.

Can I contribute GPU power?

Yes — any qualified NVIDIA or AMD single-GPU machine can join as a provider

and earn OGPU. See the provider guide above.

Where to next

• Start: openreef.network
• Providers: provider onboarding
• OpenGPU Network: opengpu.network
• Community: Telegram
• Worker source: github.com/asphyksia/openreef-finetune-worker