1. End-to-End Guide Released for Fine-Tuning Tool-Calling LLMs
A comprehensive guide has been published detailing an end-to-end supervised fine-tuning pipeline designed to improve tool-calling capabilities in small LLMs. Using the XYZ-Aquila-SFT dataset, the tutorial demonstrates how to parse multi-turn tool trajectories, extract structured tool calls, and format them into Qwen-compatible ChatML. The pipeline applies LoRA to a Qwen3-0.6B model using Hugging Face Transformers, PyTorch, and PEFT, incorporating assistant-only loss masking to focus training on the model's responses. It also features teacher-forced evaluation probes to systematically measure tool-calling accuracy improvements.
- • The tutorial outlines an end-to-end supervised fine-tuning (SFT) pipeline using Hugging Face Transformers, PyTorch, and PEFT.
- • The workflow parses multi-turn tool trajectories and converts them into Qwen-compatible ChatML format.
- • A Qwen3-0.6B model is fine-tuned using LoRA with assistant-only loss masking applied to training tokens.
- • The pipeline includes teacher-forced evaluation probes to measure tool-call performance before and after training.
Provides developers with a concrete, reproducible workflow to fine-tune small, task-specific models to reliably output structured tool calls.
2. Anthropic Details SynthID-Text Implementation for Claude
Building on its previous commitment to implement invisible watermarking, Anthropic has specified that it will utilize Google DeepMind's SynthID-Text for its future Claude models. This method adjusts token selection randomness to embed watermarks without affecting output quality, speed, or cost. The company also confirmed that images will use the C2PA standard for metadata, with a detection API planned for future release.
- • Anthropic is using Google DeepMind's SynthID-Text for text watermarking in future Claude models.
- • The method embeds watermarks by adjusting token selection randomness, ensuring no impact on quality, speed, or cost.
- • The watermark contains no user-identifying information.
- • Images will continue to use the C2PA standard for metadata.
- • A watermark detection API is planned for future release.
Provides technical clarity on the watermarking method being deployed to meet EU AI Act compliance, confirming it will not impact model performance.