AI Briefing#fine-tuning
The latest AI and developer news about #fine-tuning, with the original source and a short summary.
The latest AI and developer news about #fine-tuning, with the original source and a short summary.
The Qwen3-based CLM supports all of Jev's features and achieved 87.6% performance on Terminal-Bench.
The open-source Kev series of decision models based on Qwen3.5 has been released, offering TypeSafe API compatibility and detailed performance metrics.
A small research lab released 'Hemmingway-1', a creative writing-specialized model based on Qwen3.8-27B, under the Apache-2.0 license.
Fyxer achieved a 90% 90-day user retention rate and a 53% draft acceptance rate with an AI assistant trained on 500,000 hours of human EA data.
The Catalyst project achieved an average 27.8% performance improvement on unseen seeds by adjusting 50 hyperparameters with a single differentiation update.
Salesforce researchers proposed an On-Policy Correction pipeline to overcome the limitations of imitation learning for weaker models and facilitate the co-evolution of harnesses and models.
The OUI-1 model, fine-tuned from DiffusionGemma, achieved 71.7% on the Generative UI Benchmark, releasing high-quality UI generation technology capable of local execution.
The 'Auto-BenchMax' data synthesis pipeline, tailored to benchmark scoring methods, has been released and improved the MCP-Atlas score by more than double.
Apple released PROOF-Gen, which recovers failed tool-calling scenarios to significantly improve model distillation performance.
Sentence Transformers v6.0 introduces multi-vector model training capabilities, significantly improving domain-specific search performance.
A guide has been released to support the fine-tuning and evaluation processes of open-source and OpenAI models using LangSmith.
The 'Scaffold CoT' dataset, comprising 4 million entries, has been released to improve the reasoning accuracy of small models with 5B parameters or fewer.
TaoLive unveiled the HAT technique to train small LLMs for real-time live commerce to adapt to evolving harnesses.
Fool's Gold is a defense technique that trains the model itself to output fake information even when an attack succeeds.
CohereLabs released a 2.4B vision-language model under the Apache 2.0 license.
DLR-Lock makes unauthorized modification of pre-trained weights difficult while maintaining model performance.
With just $500 in low-cost fine-tuning, a 9B model achieved performance on par with frontier models.
Reasoning-Medical-27B, a model built on Qwen3.6-27B that maximizes medical reasoning ability, has been released.
Through low-cost RL fine-tuning, a 9B-scale open model outperformed high-performance models on specific domain performance.
A fine-tuning toolkit has been released that lets the ultra-compact TTS model Inflect-Nano-v2 be trained on a specific voice or language.
Poolside's Laguna model claims to be its own model when asked in English, but admits to being based on Qwen when asked in Chinese.
ElevenLabs has released the Finetunes Music API, enabling users to fine-tune music models with their own audio.
Inkling, a 975B-scale open-weight multimodal model designed for fine-tuning, has been released.
A public leaderboard has launched to compete on LoRA fine-tuning speed under identical hardware and dataset conditions.