futo-keyboard-ptbr

A privacy-respecting, on-device keyboard language model for Brazilian Portuguese (pt-BR), built for FUTO Keyboard.

This is a small decoder-only GPT (~42M params, llama GGUF architecture) that runs entirely on-device in FUTO Keyboard for next-word prediction, with the vocabulary conventions (and optional autocorrect/swipe training data) that FUTO's runtime expects.

Files

File	Quantization	Size	Use
pt_br_keyboard_q6k.gguf	Q6_K	~40 MB	Recommended for FUTO Keyboard
pt_br_keyboard_f32.gguf	F32	~191 MB	Full precision, for further quantization

Model details

Property	Value
Architecture	llama (RMSNorm, RoPE θ=10000, SiLU-gated FFN, tied lm_head)
Parameters	~42M
Embedding dim	512
Layers / heads	8 / 8
Context length	256 tokens
Vocabulary	16,000 (SentencePiece BPE, inverted space, byte fallback)
Tokenizer	sentencepiece (embedded in keyboardlm.ext_tokenizer_data)
Features	base_v1 inverted_space xbu_char_autocorrect_v1 xc0_swipe_typing_v1 char_embed_mixing_v1

Install in FUTO Keyboard

1. Download pt_br_keyboard_q6k.gguf.
2. In FUTO Keyboard, go to Settings → Keyboard models → Import and select the .gguf file.

Training

• Data: cleaned pt-BR corpora (MADRAS1 corpus-ptbr real subset, Wikipedia pt, Pt-Corpus-Instruct) — see the source repo for the full pipeline.
• Objective: causal next-token over [BOS] + encode(line + " "), mirroring FUTO's runtime tokenization.
• Hardware: NVIDIA CUDA, AMD ROCm, or Apple MPS.

The keyboardlm.finetuning_count is 0 — this is a from-scratch model, not a fine-tune.

License

Weights are released under Apache-2.0. Note: the underlying corpora each carry their own licenses — review them before redistributing training data or derived models.