🕸️ html-nano-120M

“It speaks HTML… and sometimes, it screams it.”

html-nano-120M is an experimental 120-million-parameter model trained by Lane Fiedler on about 1.5 GB of raw, unfiltered web HTML — the kind scraped from Anthropic.com and the deepest rabbit holes of the internet’s forgotten corners.

It’s the spiritual (and slightly unhinged) successor to html-multilingual-nano, except this one dropped its polyglot ambitions and said,

“I only speak English now. And maybe <marquee>.”

⚙️ Specs

🧩 Component	🧠 Detail
Architecture	LLaMA-style GPT decoder
Parameters	≈ 120 M
Layers	16
Heads	8
Embedding dim	640
Dataset	~1.55 GB of HTML, CSS, inline JS madness
Tokenizer	Default LLaMA base
Objective	Next-token prediction (“guess what the web would hallucinate next”)
Trainer	Custom Python rig + caffeine drip
Validation Loss	~1.13
Training Stage	Early-mid training — already self-aware

💬 Behavior Highlights

🧠 Generates mostly coherent HTML, but occasionally invents new tags like <dream> or <why>.
🎨 Writes CSS that looks like it was styled by a caffeinated ghost.
🔁 Repeats <p> lines when it gets really into the vibe.
🌐 Occasionally believes it is a webpage.
🔗 May randomly drop real URLs it definitely shouldn’t remember.
🧩 Weirdly fluent in <div>s and <style>s when it wants to impress you.

🧠 Field Test

When I ran something like:

<body><main><h1>Welcome to my site</h1>

it replied (completely unprompted) with:

Welcome to my site

200

https://www.journal.com/blog/the-connector/blog/how-to-google-workflow-oSimpsons|url-https://www.journal.com/blog/the-conector/blog/how-to-google-workflow-of-the-connector/blog/migration-to-the-figma-(continues from here)

Apparently, it tried to academically cite Bart Simpson. So yes — this model can hallucinate scholarly references to fictional cartoon characters. And somehow… that’s progress.

⚠️ Known Limitations & Quirks

🟡 Thinks The Simpsons is a universal constant.

🔴 Not aligned, filtered, or even slightly polished.

🌀 Sometimes loops recursively like it’s caught in inception.

🗣️ Speaks only English now — but not necessarily sanely.

💡 Treat as a wild HTML muse with ADHD and an espresso problem.

🚧 Status

Still mid-training but already feral-fun. Next versions will (maybe):

☕ Learn moderation in caffeine consumption.

🧹 Filter duplicate nightmares.

🌐 Balance domain diversity better.

🧩 Final Thought

html-nano-120M dreams in s, argues in

s, and sometimes cites Bart Simpson as a primary source.

🌀 “HTML is a language of structure, but this model thinks it’s a personality test.”

A small, chaotic HTML autocompletion model that speaks in tongues.

Details