AI / ML

This paper addresses the high inference costs of cross-encoders in RAG pipelines by fine-tuning LLaMA 3 (8B) as a drop-in reranker. The process involves supervised fine-tuning with LoRA adapters using the Unsloth framework, followed by 4-bit quantization to optimize performance. Evaluated on a domain-specific QA benchmark, the model significantly improves answer relevancy, context precision, and answer accuracy compared to standard retrievers.

• Replaces expensive cross-encoders with a quantized 8B LLM for real-time RAG reranking.
• Achieves 14% gain in answer relevancy and 16% in context precision via distillation.
• Uses Unsloth and LoRA for efficient fine-tuning, then 4-bit quantization for deployment.
• Validated on domain-specific benchmarks using RAGAS framework for rigorous metrics.

Researchers propose TAKE, a framework that distills large-scale text corpora down to 0.1% of their original size while maintaining downstream task fidelity. The method leverages influence functions to quantify each sample's contribution to the training objective. By convolving knowledge-based influence along the training trajectory, TAKE generates a single per-sample score to identify and retain the most informative data points.

• Reduces training data volume to 0.1% of original size, significantly lowering storage and compute costs.
• Uses influence functions to measure individual sample contribution to downstream objectives.
• Convolve influence along training trajectory to capture dynamic informative samples.
• Enables efficient continual learning by retaining only high-impact text data.

This research paper introduces a framework that decomposes image-based retinal diagnoses into structured argumentation components based on the Toulmin model. Instead of treating ML predictions as final claims, the system extracts biomarkers as grounds, uses a MedGemma agent to validate the warrant, and defines qualifiers and rebuttals. This approach shifts from black-box outputs to a structured, interpretable diagnostic assistance model.

• Decomposes ML predictions into Toulmin's claim, grounds, warrant, qualifier, rebuttal, and backing.
• Uses specialized models for biomarker extraction (grounds) rather than raw image features.
• Employs a MedGemma agent to analyze the warrant linking evidence to the diagnostic claim.
• Enables structured interpretability by explicitly modeling qualifiers and potential rebuttals.

SupplyNetPy is a new open-source Python library designed for modeling and simulating arbitrary multi-echelon supply chain networks. It supports discrete-event simulation with features like perishable inventory, node disruptions, and stochastic demand. The tool allows users to define networks as graphs and provides detailed performance reports, validated against analytical benchmarks and commercial tools.

• Open-source Python library for high-fidelity supply chain simulation
• Supports arbitrary multi-echelon structures and discrete-event modeling
• Handles perishable inventory, disruptions, and stochastic lead times
• Validated against commercial tools and analytical benchmarks
• Extensible architecture via inheritance for custom components

AWS has added OpenAI's bidirectional privacy-filter model to SageMaker JumpStart, enabling one-click deployment for PII detection and masking. The model performs context-aware token classification in a single forward pass, identifying spans like emails, names, and secrets. This allows engineering teams to build high-throughput data sanitization workflows directly on AWS infrastructure.

• OpenAI's bidirectional model now available via SageMaker JumpStart for easy deployment.
• Detects PII spans including emails, names, secrets, and account numbers in one pass.
• Designed for high-throughput sanitization workflows on AWS or on-premises.
• Tunable architecture supports context-aware masking for sensitive text data.

AWS has added Qwen3-VL-Embedding-2B and Qwen3-Reranker-4B to SageMaker JumpStart. The embedding model handles initial recall for mixed modalities like text, images, and video. The reranker then refines these results for higher precision retrieval.

• Qwen3-VL-Embedding-2B supports text, images, screenshots, and video inputs for cross-modal search.
• Qwen3-Reranker-4B improves precision by re-ranking initial recall results from the embedding model.
• Both models are available via SageMaker JumpStart for immediate deployment on AWS infrastructure.
• This pairing enables building comprehensive, multi-stage retrieval pipelines with specialized capabilities.

AWS has added Google DeepMind's Gemma-4-E2B-it to SageMaker JumpStart, offering a multimodal model optimized for local execution. The model supports text, image, and audio inputs with native function calling and a step-by-step reasoning mode. It enables agentic workflows and handles complex tasks like OCR, code generation, and UI understanding.

• Gemma-4-E2B-it is available via SageMaker JumpStart for rapid deployment
• Model optimized for efficient local execution on AWS infrastructure
• Supports multimodal inputs including text, image, audio, and video
• Includes built-in reasoning mode for step-by-step problem solving
• Enables agentic workflows through native function calling capabilities