# Mangaba AI: brier

> A Portuguese typed-decision model by Dheiver Santos (Mangaba AI): a rank-16 LoRA on 4-bit Qwen3-4B-Instruct-2507 that answers choice, score and noul questions in one pass by reading option-label logits, with a conformal answer set. v3 is v2 retrained against prompt injection.

- Page: https://systemonemodels.ai/mangaba-ai/brier
- API: https://api.systemonemodels.ai/v1/models/mangaba-ai/brier
- Download: `pip install systemonemodels && systemone pull mangaba-ai/brier`

## Facts

| | |
|---|---|
| Maker | Mangaba AI (https://systemonemodels.ai/mangaba-ai) |
| Decides | choice, score, noul |
| Architecture | brier |
| Base model | qwen/qwen3-4b-instruct-2507 |
| Parameters | 4.0B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-10-05 |
| Latest version | 3.0.0 |

## Reported evaluation

Suite: Mean over three held-out human-labelled Portuguese sets (ASSIN2, B2W, tweetSentBR; 250 items each), the maker's own evaluation. Numbers are the publisher's own.

- Decision accuracy: 67.3%

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Brier v3

A Portuguese typed-decision model by Dheiver Santos (Mangaba AI): a rank-16 LoRA on 4-bit Qwen3-4B-Instruct-2507 that answers choice, score and noul questions in one pass by reading option-label logits, with a conformal answer set. v3 is v2 retrained against prompt injection.

A personal project of Dheiver Francisco Santos, published as Mangaba AI. Portuguese only. The repo holds the MLX LoRA adapter (73 MB), not the base; the brier package dequantises the 4-bit base and merges the adapter in PyTorch off Apple Silicon (CPU is very slow). A block attention mask answers all questions in one pass. Trained against soft majority-vote targets from LLM labellers; conformal (APS) answer sets ship with a calibration file, and calibration is reported as Brier score only, with no ECE. On the maker's held-out human-labelled Portuguese sets v3 averages 67.3% (v2: 70.1%, more accurate on clean text); prompt-injection attacks succeed in 16.7% of the maker's cases; accuracy drops on texts over about 2,000 tokens. The maker compares it with TypeSafe Jev and says no Brier version beats Jev yet. The package's "rapido" mode runs Convai's Laya weights unchanged and is not part of this entry. Serves /v1/decide and a /v1/systemone migration route.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 4B |
| Base model | `qwen/qwen3-4b-instruct-2507` |
| Maker | Mangaba AI |
| Released | 2026-10-05 |
| License | apache-2.0 |
| Reported accuracy | 67.3% |
| Reported latency | about 1.3 s median per request on a MacBook Air M5 (MLX) |

## Get the weights

```bash
pip install systemonemodels
systemone pull mangaba-ai/brier
```

The files are served from the maker's Hugging Face repository, [`mangaba-ai/brier-v3`](https://huggingface.co/mangaba-ai/brier-v3), and verified against the checksums recorded here.

## Read more

- [Model card (v3)](https://huggingface.co/mangaba-ai/brier-v3)
- [Brier v2](https://huggingface.co/mangaba-ai/brier-v2)
- [Code, data and evaluation](https://github.com/Mangaba-ai/brier)

---

*This page was opened by System One for Mangaba AI, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models