Assaoka/Tucano2-qwen-0.5b-Merge-ReLiSA

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Assaoka/Tucano2-qwen-0.5b-Merge-ReLiSA is an 0.8 billion parameter Qwen-based model developed by Assaoka, specifically designed for sentiment analysis in Portuguese. This model specializes in classifying sentiment (positive, negative, neutral, mixed) within literary reviews, leveraging a unique Knowledge Accumulation (Prior) strategy. It was fine-tuned on the ruanchaves/reli dataset after merging several expert models, demonstrating significant performance improvements over its baseline in in-domain tasks.

Loading preview...

Tucano2-qwen-0.5b-Merge-ReLiSA: Portuguese Sentiment Analysis Expert

This model, developed by Assaoka, is a specialized 0.8 billion parameter Qwen-based language model focused on sentiment analysis in Portuguese. It was created using a Knowledge Accumulation (Prior) strategy, involving an initial merge of several expert models (Brighter, FinBERT, Phrasebank, GoEmotions) followed by fine-tuning on the ruanchaves/reli dataset.

Key Capabilities and Features

  • Task: Sentiment analysis (positive, negative, neutral, mixed) for literary reviews.
  • Domain: Specifically optimized for Portuguese (PT-BR) language content.
  • Development Strategy: Utilizes sequential alignment via Knowledge Accumulation (Prior) and subsequent fine-tuning.
  • Architecture: Built upon the Qwen 0.5B base, enhanced through a TIES Merge of multiple specialized models.

Performance Highlights

Evaluation metrics show strong performance, particularly in its target domain. For the RELI-SA dataset, the model achieved a Macro F1 of 69.75% and an Accuracy of 80.41%, significantly outperforming the baseline model (Macro F1 30.66%, Accuracy 37.07%). While excelling in-domain, performance on out-of-domain datasets like FINBERT-PT-BR and FINANCIAL-PHRASEBANK shows varying results, indicating its specialization.

Good for

  • Developers needing a dedicated and highly accurate model for sentiment analysis of literary reviews in Portuguese.
  • Applications requiring structured JSON output for sentiment classification.
  • Research into model merging and knowledge accumulation techniques for specialized NLP tasks.

Finetunes
1 models