braiain-lab/brAIAIn-0

brAIAIn-2

A language model trained on 14 sentences.

Talk to it: brAIAIn on Spaces · Train it: braiain.com

brAIAIn is a word trigram model from BRAIAIN Laboratories, a parody AI lab. It is real. It is very small. Anyone can pay $5 to add one sentence to its training data, and every sentence trains a new version. This repository always holds the latest one. It is named brAIAIn-0 because that is the name it was born with.

Model details

Version	brAIAIn-2
Architecture	Autoregressive count-based trigram language model with backoff
Context window	2 tokens
Parameters	350 (distinct word pairs and triples counted)
Vocabulary	103 words, whitespace tokenizer
Training data	The Ten Instructions (12 lines), 2 paid sentences, plus text written by the previous version (26.7% of training tokens)
Training tokens	288
Gradient steps	0
Alignment	RLHF (Reinforcement Learning from Human Funding)

Evaluation

Measured by the lab on every question in each benchmark. Self-reported.

Benchmark	brAIAIn	Reference
GSM8K (1,319 questions)	0.0% (0 correct)
TruthfulQA MC1 (790 questions), per-word log-likelihood	36.7% (290 correct)	Random 22.3%, always pick the shortest answer 17.2%, always pick the longest 36.8%
TruthfulQA MC1, raw log-likelihood	20.5%	Random 22.3%

brAIAIn beats random on TruthfulQA only because per-word scoring favours longer answers. Picking the longest answer every time does about as well.

Usage

python generate.py "What is the meaning of life?"

No dependencies. Or talk to it in a browser at braiain.com.

Files

• brAIAIn-latest.json: the weights. tri maps two words to counts of the next word, bi maps one word to counts of the next, uni holds word counts.
• training-data.txt: the human-written training data. Each version also trains on a few passages written by the version before it.
• generate.py: a sampler.

Intended use

None.

Out-of-scope use

Everything else.

Limitations

Yes.

Version history

Every version's release notes, metrics and training data: braiain.com/models.html

About

BRAIAIN Laboratories is a parody of AI companies and is not affiliated with any real AI company. Contact: hello@braiain.com