Sedibaai/SedibaLM

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SedibaLM V7 is a 1.54 billion parameter Sepedi (Sesotho sa Leboa) language model developed by Sediba AI NPC, fine-tuned from Qwen2.5-1.5B-Instruct. It is specifically optimized to treat Sepedi as a native language, addressing the under-representation of this South African language in existing models. This model excels in Sepedi conversational AI and community information delivery, demonstrating sovereign African AI capability.

Loading preview...

SedibaLM V7: A Sepedi-Optimized Language Model

SedibaLM V7, developed by Sediba AI NPC, is a 1.54 billion parameter language model fine-tuned from Qwen2.5-1.5B-Instruct using QLoRA. Its primary goal is to provide robust support for Sepedi (Sesotho sa Leboa), a South African language spoken by approximately 4.7 million people, which is severely under-represented in current LLMs.

Key Capabilities & Differentiators

  • Native Sepedi Understanding: Unlike its base model, V7 processes Sepedi input directly, providing responses in Sepedi rather than defaulting to English or misidentifying the language.
  • Improved Monolinguality: Evaluation shows a nearly 3x improvement in Sepedi monolinguality compared to the base Qwen2.5-1.5B model, indicating its strong focus on generating Sepedi content.
  • ChatML Format: V7 is a ChatML-format SFT model, requiring the use of a chat template for proper interaction to avoid immediate End-Of-Sequence (EOS) tokens.
  • Resource-Efficient: Built on a 1.5B parameter base, it offers a specialized solution for low-resource language modeling research and deployment.

Intended Use Cases

  • Sepedi Conversational AI: Ideal for building chatbots and interactive systems in Sepedi.
  • Community Information Delivery: Facilitating access to information for Sepedi-speaking communities.
  • Low-Resource Language Modelling Research: A valuable tool for advancing research in under-represented languages.
  • Demonstrating Sovereign African AI: Showcasing local AI development capabilities for South African languages.

Limitations

It's important to note that V7 is not aligned (no RLHF or DPO), has modest continuation chrF, and can still exhibit code-switching. Its 1.5B parameter size means it's not comparable to frontier models for complex reasoning. The model weights are licensed under CC BY-SA 4.0, with training data governed by a separate NOODL license.

Finetunes
1 models