---
license: mit
datasets:
- SupraLabs/chat-titles-filtered-115K
- ogrnz/chat-titles
- Michionlion/chat-titles-english
language:
- en
---
# TinyTitle

A tiny model (~1.8M params) that turns a chat message into a short title (2-10 words). English only. The whole thing (model + tokenizer + runtime) runs in under 5 MiB of ram, in a few tens of ms, on one small C binary.

It's a small GRU with a copy trick: for each title word it either makes up a word from its 8k vocabulary or copies a whole word straight from your message (so it keeps your spelling and casing).

Honest note: an LLM wrote almost all of this while I nodded along. It works tho. It's a toy, not a real llm.

## quickstart

This is a custom format (`.ttm1`) with a custom C runtime, so it does not work with transformers or llama.cpp. You need the runtime source from the [github repo](https://github.com/azomDev/TinyTitle).

```bash
# build the runtime (needs cc, nothing else)
cc -std=c11 -O3 -DNDEBUG -o title-v1 runtime/main.c -lm

# run it
./title-v1 model.ttm1 "Why does my wifi keep dropping?"
```

Files in this repo:

- `model.ttm1` - the int8 model (1.98 MB)
- `tokenizer-8k.json` + `tok-8k.ttok` - the tokenizer
- `runtime/` - main.c, ttm.h, Makefile (the whole runtime is just those files)

## numbers

| metric | value |
|---|---|
| params | ~1.8M (int8) |
| model file | 1.98 MB |
| peak rss | 4.89 mib (all pages touched) |
| typical cpu | ~25 ms |
| vocab | 8k unigram |

## train

Training needs python + torch. Full instructions in the [github repo](https://github.com/azomDev/TinyTitle). It trains in under 0.5 GB of vram.

## license

MIT. Training data comes from three public huggingface datasets ([SupraLabs/chat-titles-filtered-115K](https://huggingface.co/datasets/SupraLabs/chat-titles-filtered-115K) cc-by-4.0, [ogrnz/chat-titles](https://huggingface.co/datasets/ogrnz/chat-titles) MIT, [Michionlion/chat-titles-english](https://huggingface.co/datasets/Michionlion/chat-titles-english) cc-by-4.0), assembled by the github repo's `build_dataset.py`.