MiniCPM5-1B Viva Mais v4 GGUF
This repository contains the llama.cpp-compatible GGUF export of marinarosa/minicpm5-1b-vivamais-v4.
- Quantization:
Q4_K_M - File:
minicpm5-1b-vivamais-v4-Q4_K_M.gguf - Source model:
marinarosa/minicpm5-1b-vivamais-v4 - Training dataset:
marinarosa/minicpm5-vivamais-text-sft-v4 - Fine-tuning base:
marinarosa/minicpm5-1b-vivamais-v1 - Upstream family:
openbmb/MiniCPM5-1B
Source Model Summary
The source model is a Brazilian Portuguese Viva Mais text model for grounded Q&A over travel-agency WhatsApp/CRM context. It was trained with response-only SFT on 4000 redacted rows, using 4-bit base loading with LoRA rank 16, alpha 32, and A100 training. The final source artifact was merged to 16-bit before GGUF export.
Eval Results
| Suite | v1 baseline avg | v4 source avg | v4 pass rate |
|---|---|---|---|
| Viva Mais dashboard QA | 0.5042 | 0.5570 | 0.4494 |
| Generic PT-BR conversational QA | 0.5814 | 0.6279 | 0.6279 |
Known limitation: the model may over-generate after the direct answer and include adjacent customer facts. Use short generation limits and isolated retrieval contexts.
llama.cpp Usage
llama-cli \
--model minicpm5-1b-vivamais-v4-Q4_K_M.gguf \
--ctx-size 8192 \
--temp 0 \
--prompt "<s><|im_start|>system\nResponda em PT-BR usando apenas o contexto fornecido.<|im_end|>\n<|im_start|>user\nContexto: ...\nPergunta: qual saldo falta?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
Privacy
The source training data is synthetic or redacted. No raw WhatsApp exports, identity documents, client names, CPF values, phone numbers, emails, card numbers, or local private data are included in the published dataset or model.