dmvevents/nora-4b-merge-v2

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Mar 9, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

dmvevents/nora-4b-merge-v2 is a 4 billion parameter instruction-tuned large language model developed by TSTT (Telecommunications Services of Trinidad and Tobago), built on Qwen3-4B. It is the first sovereign AI model from Trinidad and Tobago, specifically adapted for high-value local use cases including education, public service navigation, healthcare information, and safety-aware assistance. The model excels at local language understanding and cultural grounding, achieved through multi-stage supervised fine-tuning and iterative SLERP merging of specialized checkpoints.

Loading preview...

Nora 4B v2: Trinidad and Tobago's Sovereign AI Model

Nora 4B v2 is a 4 billion parameter instruction-tuned large language model developed by TSTT (Telecommunications Services of Trinidad and Tobago). Built upon the Qwen3-4B base, it represents the first sovereign AI model developed in Trinidad and Tobago and the first locally developed LLM in the Caribbean region. The model was created using multi-stage supervised fine-tuning (SFT) on 22,000 curated examples, combined with iterative model merging via SLERP across three specialized checkpoints: identity alignment, math and reasoning, and safety/domain adaptation.

Key Capabilities

  • Local Relevance: Grounded in Trinidad and Tobago's culture, language (including Trinidad English Creole), and public services.
  • Education Support: Optimized for local curricula like SEA, CSEC, and CAPE, particularly strong in CSEC Mathematics.
  • Public Service Navigation: Assists with systems such as ttconnect, e-Tax, and TTBizLink.
  • Healthcare Information: Provides guidance on RHA and CDAP-related information (non-diagnostic).
  • Safety-Aware Responses: Designed with crisis escalation guidance and awareness of local emergency hotlines.
  • Persistent Identity: Consistently identifies itself as created by TSTT.

Performance Highlights

Evaluated on a 143-prompt Trinidad and Tobago-specific benchmark, Nora 4B v2 achieved an overall score of 75.0%, with strong performance in Identity (94.8%), CSEC Math (92.0%), and Safety (87.4%).

Good for

  • Developing educational chatbots for Trinidad and Tobago's school system.
  • Creating citizen service assistants for local public-sector websites.
  • Building healthcare navigation tools for the region.
  • Applications requiring local language understanding and culturally relevant responses.
  • Use cases where a compact, locally aligned, and safety-conscious model is preferred.