bard
A small GPT trained on Shakespeare, stored onchain. Weights stream from a contract on Ethereum Sepolia, the artifact hash is verified in the browser, and inference runs in a Web Worker.
bard has no database and no JSON API. It is a static page that reads the model directly from the chain, so it sits apart from the content services.
How it works
The weights are chunked, gzipped, and written into the contract's storage. The browser reassembles the model client-side in four stages:
- Pointers. Read the chunk pointers from the contract.
- Chunks. Fetch each weight chunk over a public Sepolia RPC.
- Decompress. Gunzip the reassembled artifact.
- Verify. Recompute the sha-256 hash and confirm it matches the artifact hash the contract advertises.
The model only loads when the in-browser hash matches the onchain one. This is the same content-addressing the rest of Farfield uses, enforced in the browser against the chain.
Contract
Address 0xBA2D82930b2F74B1319Fd326bdF43b567Ac03720 on Ethereum Sepolia (chain 11155111). Inspect it on Etherscan or Sourcify.
Inference
Generation runs in a Web Worker so the page stays responsive. The sampling controls are the usual set: temperature, top-k, top-p, max tokens, frequency penalty, and a seed. Output streams token by token. There is no wallet and no server-side inference; everything after the chunk fetch happens locally.