PleIAs/Cassandre-RAG

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 17, 2024Architecture:Transformer0.0K Featherless Exclusive Cold

PleIAs/Cassandre-RAG is an 8 billion parameter Llama-3.1-based model fine-tuned for Retrieval Augmented Generation (RAG) on French administrative documents, specifically focusing on school administration sources. It is designed to efficiently perform RAG tasks by expecting a predefined input structure and clearly citing specific excerpts and source documents in its generated answers. The model was trained using synthetic queries, BM25-retrieved documents, and generated responses to enhance its performance in this specialized domain. Its 32768 token context length supports processing substantial administrative texts.

Loading preview...

Cassandre-RAG: Specialized RAG for French Administrative Documents

Cassandre-RAG is an 8 billion parameter model built upon the Llama-3.1 architecture, specifically fine-tuned for Retrieval Augmented Generation (RAG) tasks. Its primary focus is on processing and generating responses based on French administrative documents, with a particular emphasis on school administration sources.

Key Capabilities

  • Efficient RAG: Designed to handle RAG tasks very efficiently by adhering to a predefined input structure.
  • Source Citation: Generates answers that clearly cite specific excerpts and the source documents used, ensuring traceability and accuracy.
  • Specialized Training: Fine-tuned on a unique corpus comprising synthetic queries, documents retrieved via BM25, and generated answers, all derived from French administrative texts.
  • Structured Input/Output: Expects a specific prompt format including "Query," "Source," and "Answer" sections, and outputs answers with embedded references using a <ref text="[Quoted text from source]">[Source ID]</ref> format.

Training Details

The model underwent 3000 training steps with a learning rate of 3e-4 and a maximum sequence length of 8192. It utilized LoRA configuration with an alpha of 16, dropout of 0.1, and R of 64, targeting key attention and feed-forward modules. Quantization was applied using 4-bit nf4 type with float16 compute dtype.

Good For

  • Applications requiring precise information retrieval and generation from French administrative documents.
  • Systems needing verifiable answers with clear source attribution.
  • Developers building RAG solutions for educational or public sector administrative contexts in France.