marinarosa/minicpm5-vivamais-text-sft-v4

MiniCPM5 Viva Mais text SFT v4

This dataset contains the redacted training and evaluation artifacts used for marinarosa/minicpm5-1b-vivamais-v4. It was built for Viva Mais, a local-first Portuguese WhatsApp travel-agency copilot that answers grounded questions from an extracted CRM context.

Files

• —data/train.jsonl: 4000 chat-format SFT rows.
• —data/eval/vivamais_qa_eval.jsonl: 158 dashboard QA eval rows.
• —data/teacher/rio31_teacher_distill.jsonl: 80 accepted rows distilled from prefeitura-rio/Rio-3.1-Open-4B-Instruct.
• —metadata.json: source volume, row counts, training recipe, and eval summary.

Each row is JSONL. Training rows use a messages list with chat roles. Eval rows contain redacted Viva Mais dashboard contexts, questions, and scoring metadata.

Dataset Mix

The v4 recipe deliberately emphasized grounded refusal and entity isolation over open-ended chat style. The target weights were negative 0.60, domain QA 0.38, teacher distillation 0.02, and generic/public/travel-vocab buckets 0.0. A small 28-row generic remainder was still included by the dataset builder's fill logic.

Teacher Distillation

• —Teacher model: prefeitura-rio/Rio-3.1-Open-4B-Instruct
• —Accepted teacher rows: 80
• —Rejected teacher rows: 5
• —Prompting goal: final-answer-only Brazilian Portuguese responses, avoiding hidden-reasoning wrappers and verbose assistant preambles.
• —Filtering: removed rows with wrapper phrases, hidden reasoning artifacts, weak/unsafe travel claims, and answers not matching the allowed v4 categories.

Privacy

This dataset is synthetic or redacted. Customer names, phone numbers, emails, CPF values, card numbers, transaction identifiers, hotels, and localizers use placeholders such as [CLIENTE_1], [TELEFONE_1], [HOTEL_1], and [LOCALIZADOR_1]. No raw WhatsApp exports, identity documents, client names, full private transcriptions, or local private data are included.

Evaluation Summary

The model trained from this dataset improved over marinarosa/minicpm5-1b-vivamais-v1 on both eval suites but did not pass the internal publish gate because leakage failures stayed high.

Known weakness: the candidate often starts with the correct answer but continues generating unrelated customer facts. This causes cross-customer leakage failures.