AI / ML

Amazon SageMaker AI extends its serverless model customization to include full fine-tuning capabilities for over 25 open-source models, including Llama, Gemma, and Qwen families. This update allows engineers to update all model parameters rather than relying solely on parameter-efficient methods like LoRA. The feature enables deeper adaptation for domain-specific patterns, specialized reasoning, and complex output formats using proprietary datasets.

• Full fine-tuning is now available in SageMaker serverless, replacing the need for dedicated GPU instances for this workload.
• Supports 25+ open-source models including Llama, Gemma, Qwen, Nemotron, and gpt-oss families.
• Enables deep adaptation for specialized reasoning and complex output structures beyond surface-level style adjustments.
• Ideal for internalizing domain knowledge from large proprietary datasets where LoRA may be insufficient.
• Simplifies infrastructure by removing the operational overhead of managing persistent fine-tuning clusters.

Firecrawl has open-sourced a Rust library that classifies PDFs as text-based, scanned, or mixed in under 50 milliseconds. The tool extracts text with position awareness and converts documents to Markdown without relying on expensive OCR services. It includes bindings for Python, Node.js, and WebAssembly to facilitate local processing.

• Skip OCR costs for the ~54% of PDFs that are text-based by using smart content stream sampling.
• Get per-page OCR routing decisions with confidence scores to handle mixed or scanned documents selectively.
• Deploy locally via Python, Node.js, or WebAssembly bindings for low-latency processing under 200ms.
• Extract text with position awareness to preserve document structure when converting to Markdown.

Nova is an automated end-to-end JIT compiler built on MLIR designed to overcome the limitations of eager execution in high-level tensor frameworks. It captures eager executions to provide whole-graph visibility, enabling granular control over hardware mapping, operation fusion, and memory hierarchy optimization. The system tunes execution down to the register level to maximize physical hardware utilization for deep learning models at scale.

• Bridges the gap between flexible model design and low-level hardware optimization via whole-graph analysis.
• Captures eager executions to enable cross-operation fusion and register-level tuning.
• Provides granular control over memory hierarchies to maximize physical hardware utilization.
• Leverages MLIR for an automated end-to-end compilation pipeline from high-level ops to hardware.

The C-Guard framework addresses the conflict between catching real harm and avoiding over-refusal in RL alignment by using a constitution-grid instrument to generate training data. It employs C-LIM, a per-cell learnability score, to dynamically prune, densify, amend, or expand data regions based on their utility. This approach identifies and eliminates dead-weight data, such as 187 untargeted rows that provided zero gain, while significantly improving learning impact in those regions.

• C-Guard resolves the safety vs. utility conflict in RL alignment through structured data generation.
• C-LIM scores per-cell learnability to optimize data composition before training begins.
• Method eliminated 187 zero-gain rows, lifting learning impact from 0.733 to 0.80.
• Reduces over-refusal rates from 22.4% to 12.8% without worsening adversarial vulnerability.