Heda

Label Studio, but onchain. Annotate datasets, earn crypto, fine-tune AI models all on the 0G stack.

What is Heda?

Heda is a decentralized data annotation marketplace where:

• Job Creators upload raw data (images or text), post annotation jobs with ETH bounties

• Annotators pick up jobs, label data, and get paid instantly onchain per approved task

• Developers purchase verified datasets and fine-tune LLMs on 0G Compute

Every piece of data lives on 0G Storage (permanent, verifiable). Every payment is settled on 0G Chain (trustless, instant). Every fine-tune runs on 0G Compute (decentralized GPU).

0G Stack Usage

Architecture

Raw workflow

• Creator uploads data → 0G Storage (root hash)

• Creator posts job → AnnotationMarket.sol (bounty locked onchain)

• Annotator claims task → claimTask() (30-min reservation, prevents wasted work)

• Annotator annotates → Workspace UI (bbox/polygon for images, classification for text)

• Annotator submits batch → submitBatch() (1 MetaMask signature for all tasks)

• Creator approves → approveWork() (instant ETH payment, auto-closes when complete)

• Creator publishes → COCO JSON (images) or JSONL (text) → 0G Storage

• Developer purchases → DatasetRegistry.sol (royalty to creator)

• Developer downloads → ZIP with images + annotations (image) or JSONL (text)

• Developer fine-tunes → 0G Compute Router API (text datasets only)

Deployed Contracts (Galileo Testnet)

Annotation

- Image annotation : bounding boxes (draw, drag, resize with handles), polygons

- Text annotation : single-label classification with configurable labels

- Batch submit : annotate all tasks, sign once (1 MetaMask popup for N tasks)

- Task claiming : 30-min reservation prevents multiple annotators wasting work on same task

- Draft saving : annotations saved to localStorage, restored on page reload

- Preview before submit : review all tasks (done/pending) before signin

Jobs

- Auto-close : jobs automatically close when all tasks are approved

- Multi-annotator support : each task claimed independently, parallel work on different tasks

- Instant payment : annotator receives ETH immediately on approval

Datasets

- COCO JSON export image datasets: standard COCO format with images, annotations, categories

- JSONL export text datasets: 3 schemas supported (Chat Messages, Instruction, Text Completion)

- ZIP download image datasets download as ZIP with images/ folder + annotations/instances.json

- Onchain provenance every dataset root hash links to source data and annotations on 0G Storage

- Royalties purchase price goes directly to publisher (no platform cut)

Fine-Tuning (Text only)

- 0G Compute integration fetches JSONL from 0G Storage, uploads to 0G Compute, polls status

- Supported models Qwen2.5-0.5B-Instruct ($0.5/M tokens), Qwen3-32B ($4/M tokens)

- Schema-aware JSONL output matches selected schema (chat/instruction/completion)

Supported Data Formats