Write I-Lang operation chains, one per line, and read the IML beside them; edit the IML and the I-Lang follows. The conversion is exact: whatever converts comes back unchanged. Chains only: documents with declarations are converted by the reference codec.
| Text | cl100k_base tokens | UTF-8 bytes | Characters |
|---|---|---|---|
| I-Lang | … | 0 | 0 |
| IML | … | 0 | 0 |
Every IML message starts with its version and registry digest, #iml/0.5/7e29fae7f5ea, which is 17 cl100k_base tokens. For a single chain the IML message is therefore longer than the I-Lang line; the document form carries the header once for all chains. Measured figures for the full corpora are in the IML measurement reports.
Paste an I-Lang document, such as an agent's SOUL file, and read what the canon grammar validator reports. A text whose first line is ::ILANG::v5.0 is read as a raw I-Lang document; any other text is read as Markdown, and its I-Lang lines and fenced blocks are checked.
- Nothing to check yet.
Put any two texts side by side, for example an instruction written in prose and the same instruction in I-Lang, and compare their sizes. Nothing is translated here: you supply both texts.
| Text | cl100k_base tokens | UTF-8 bytes | Characters |
|---|---|---|---|
| A | … | 0 | 0 |
| B | … | 0 | 0 |
| B minus A | … | 0 | 0 |
Tokens are counted with cl100k_base, the tokenizer of GPT-4 and GPT-3.5, loaded once from this site (about 1.7 MB). Other models use other tokenizers and count differently.
Your input never leaves your browser. The converter, the validator and the tokenizer run in this page, and nothing you type is sent anywhere or stored. The tools and the tokenizer's vocabulary are loaded from ilang.ai; the site's Cloudflare Web Analytics counts page views, not what you type.