mcit/fanar-gov-7b-instruct
Fanar Gov 7B Instruct
Grounded QA F1 (gov eval)
0.87
Citation accuracy
0.90
Arabic MMLU (subset)
0.64
Hallucination rate (grounded)
3.9%
Tokens/sec (A100, bf16)
48
About this model
Sovereign 7-billion-parameter Arabic-first instruction model tuned for government assistant and retrieval-augmented workloads. Optimised for grounded question answering over ministerial circulars, service catalogues, and legal texts, with citation-style answer formatting built into the chat template.
Intended use
Backbone for government chat assistants, RAG pipelines over official document stores, and drafting support inside entity workspaces. Answers should be grounded via retrieval for factual use cases.
Training data lineage
Instruction-tuned on curated pairs derived from the arabic-gov-docs-corpus and gov-services-catalog, on top of a sovereign Arabic-English pretraining mix.
Limitations & bias notes
Ungrounded factual recall about post-2025 regulations is unreliable and the model should always be paired with retrieval for policy questions. Long multi-turn sessions beyond 8K context show degraded Arabic diacritic consistency.
Evaluation metrics
| Grounded QA F1 (gov eval) | 0.87 |
| Citation accuracy | 0.90 |
| Arabic MMLU (subset) | 0.64 |
| Hallucination rate (grounded) | 3.9% |
| Tokens/sec (A100, bf16) | 48 |
Try it
Live sandboxFeedback
5.0(1)Owning entity
- Updated
- 2026-07-21
- Latest version
- 1.3.0
- License
- Open
- Access
- Open