A naukri.com initiative
VentureBeat
3w
224
Image Credit: VentureBeat
Ethically trained AI startup Pleias releases new small reasoning models optimized for RAG with built-in citations
- French AI startup Pleias has released two small reasoning models optimized for retrieval-augmented generation, citation synthesis, and structured multilingual output.
- The models, Pleias-RAG-350M and Pleias-RAG-1B, are based on the Pleias 1.0 family of language models and available in CPU-optimized GGUF format.
- These models aim to provide cost-effective alternatives to large-scale language models without compromising traceability, multilingual capabilities, or structured reasoning workflows.
- Pleias positions their design choice of built-in source citation as an ethical imperative that aligns with regulatory demands for explainable AI.
- The Pleias-RAG models can autonomously assess queries, determine complexity, and decide on responses based on source adequacy, offering structured and reasoned answers.
- Despite their small size, the Pleias-RAG models exhibit behavior associated with larger systems, showcasing efficient performance on standard CPUs.
- In benchmark evaluations, these models outperform larger models on tasks like HotPotQA and show strength in multilingual scenarios with minimal performance degradation.
- The models' multilingual support is achieved through careful tokenizer design and adversarial training exercises for language-switching.
- Pleias envisions their models being used to augment the performance of existing AI models in orchestration settings, highlighting their cost-effectiveness and complementarity.
- The models are released under the Apache 2.0 license, emphasizing commercial reuse and integration into various systems and applications.
Read Full Article
13 Likes
For uninterrupted reading, download the app