franciscoaleixo/SaibaTudo-Eleicao2026

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

The franciscoaleixo/SaibaTudo-Eleicao2026 is a 1.5 billion parameter QLoRA fine-tune of Qwen2.5-1.5B-Instruct, designed for Natural Language Understanding (NLU) in the context of the 2026 Brazilian General Elections. It interprets voter questions, extracting intent and entities (e.g., office, state, party, name, topic) and outputs a short JSON. This model is optimized to provide structured NLU output for an election information application, rather than generating factual responses directly.

Loading preview...

Overview

This model, franciscoaleixo/SaibaTudo-Eleicao2026, is a 1.5 billion parameter QLoRA fine-tune of Qwen2.5-1.5B-Instruct. Its primary function is to perform Natural Language Understanding (NLU) on questions related to the 2026 Brazilian General Elections. It processes user queries and outputs a concise JSON object containing the detected intent and relevant entities (such as cargo, uf, partido, nome, tema). This model is a core component of the SaibaTudo Eleições 2026 application.

Key Capabilities & Differentiators

  • Intent and Entity Extraction: Specifically trained to identify election-related intents (e.g., LISTAR_CANDIDATOS, PERFIL_CANDIDATO, PESQUISAS) and extract entities like candidate names, political offices, and states.
  • Structured JSON Output: Provides a clean, short JSON output, adhering to a strict v2 contract, which is revalidated multiple times and can be enforced with GBNF grammar in production.
  • Role as an Interpreter: Crucially, the model does not provide factual answers; it only interprets the user's question. Factual data is sourced exclusively from official TSE open data packages loaded locally on the client device.
  • Ethical Design: Explicitly designed not to recommend, compare, or predict candidates, aligning with TSE regulations. Requests of this nature are classified as RECOMENDACAO for neutral refusal.
  • Robust Training: Fine-tuned with QLoRA 4-bit NF4 on a dataset of over 20,000 examples in Brazilian Portuguese, including templates and external real-world formulations, achieving a 90.0% intent accuracy (excluding contract limitations) on evaluation gates.
  • Production-Ready: Optimized for CPU inference using GGUF Q4_K_M with llama.cpp and GBNF grammar, demonstrating low latency (0.77s average) and 0% entity hallucination.

When to Use This Model

  • NLU for Election-Related Queries: Ideal for applications requiring precise intent and entity extraction from user questions about Brazilian elections.
  • Integration with External Data Sources: Best suited for systems where factual information is retrieved from a separate, authoritative database (like official TSE data) based on the model's NLU output.
  • Building Conversational AI for Elections: Can serve as the NLU layer for chatbots or virtual assistants focused on providing structured information about the 2026 Brazilian General Elections.

Finetunes
1 models