tardellirs/aac-board-generator-770m-ptbr-GGUF

AAC Board Generator 770M (pt-BR) — GGUF

llama.cpp GGUF builds of aac-board-generator-770m-ptbr — a 771.6M Brazilian-Portuguese multi-function AAC model that drafts pictogram-board word lists and follows in-assistant board-editing instructions (add / remove / generate N), on CPU. See the model card for what it does, training, and evaluation.

Files

File	Quant	Size	Notes
aac-board-generator-770m-ptbr-Q4_K_M.gguf	Q4_K_M	561 MB	smallest; slight quality drop
aac-board-generator-770m-ptbr-Q5_K_M.gguf	Q5_K_M	606 MB	good size/quality balance
aac-board-generator-770m-ptbr-Q6_K.gguf	Q6_K	766 MB	near-lossless
aac-board-generator-770m-ptbr-Q8_0.gguf	Q8_0	824 MB	recommended — safe near-lossless
aac-board-generator-770m-ptbr-f16.gguf	f16	1.55 GB	full precision reference

At Q8_0 the model needs on the order of ~0.9–1.1 GB RAM to serve (weights + a small context) and produces a ~12-item board in about 4 s on 4 CPU threads.

Usage

llama-server -m aac-board-generator-770m-ptbr-Q8_0.gguf -t 4 -c 2048 --host 127.0.0.1 --port 8080

Gemma has no system role — put the instruction in the user turn and call /completion:

import requests
INSTR = ("Você monta pranchas de CAA (pictogramas, pt-BR). Para o PEDIDO, liste ~12 itens concretos e relevantes, "
         "um por linha, no formato palavra|tipo|sinônimos (tipo: v/s/a/e/l/p). Só a lista.")
pedido = "monta uma prancha de brincar no parquinho"
prompt = f"<start_of_turn>user\n{INSTR}\n\nPEDIDO: {pedido}<end_of_turn>\n<start_of_turn>model\n"
print(requests.post("http://127.0.0.1:8080/completion",
                    json={"prompt": prompt, "temperature": 0, "n_predict": 320}).json()["content"])

Generation is greedy / deterministic (temperature 0). Output is one item per line, word|type|synonyms.

Notes on the build (why these GGUFs are faithful)

This model is a vocabulary-trimmed Gemma 3 (262k → 64k tokens). A reconstructed SentencePiece tokenizer is not bit-exact for a trimmed vocabulary and would silently degrade greedy generation. These GGUFs instead use the BPE tokenizer.json path, which tokenizes identically to the original — so the quantized model reproduces the untrimmed parent's outputs. Verified: identical token counts and identical boards vs. the fp16 reference.

License & attribution

Released under the Gemma license (inherited from Gemma 3). Built for the ARASAAC ecosystem (CC BY-NC-SA content). See the model card for full terms.

Developed for and used in Papuguinho — www.papuguinho.com.