Today's Fine-tuning & Training: Fastest-Growing Projects — September 28, 2026
Today's the Fine-tuning & Training space on GitHub, we see a mix of projects focusing on decision-making models, dataset curation, and novel training techniques for AI agents. The most notable trend is the continued interest in fine-tuned decision engines that offer efficiency gains over traditional autoregressive models, as well as tools designed to streamline the process of working with large language models (LLMs) on personal devices.
Mapika/decider stands out this week with a growth score of 61.54 and 838 stars. It provides one-pass typed decisions with calibrated probabilities inspired by System One style models, fine-tuned from Qwen3.5-2B. The project's popularity likely stems from its innovative approach to decision-making in AI systems, offering faster and more reliable outcomes compared to conventional methods.
Liuziyu77/Valen is another prominent tool with a growth score of 48.60 and 336 stars. This repository aims to train a Jev-like multimodal model that integrates vision capabilities into decision-making processes, expanding the scope of TypeSafe Jev models. The project's rapid growth can be attributed to its unique approach to integrating visual data with AI decision engines, making it highly relevant for applications requiring both textual and image-based inputs.
AkashPriyadarshii/jev-curate has a growth score of 26.95 and 86 stars. This tool is designed to sift through high-throughput synthetic and pretraining datasets for TypeSafe Jev models, using Rust streaming core technology with Parquet and JSONL I/O formats. The project's growing popularity can be linked to its efficient dataset processing capabilities that significantly enhance the training efficiency of Jev models.
davefano/omarchy-trackpad-plus has a growth score of 13.31 and 126 stars, focusing on fine-grained trackpad controls for Omarchy devices. While not directly related to AI model training, it shows promise in enhancing user interaction with AI-driven systems through precise device customization.
friday-memory/friday, with a growth score of 11.46 and 62 stars, introduces an open-source persistent cognitive memory layer designed to prevent "amnesia" in AI coding agents like Cursor, Claude & Copilot. This project's rising star count suggests growing interest in maintaining context awareness in AI-driven development environments.
luigisaetta/llm-fine-tuning-on-mac has a growth score of 5.56 and 21 stars, offering fine-tuning capabilities for small LLMs using LoRA on Mac devices. The project's modest but steady growth indicates its relevance to developers looking to perform model tuning locally without high-end hardware requirements.
intikhab49/open-jev-typed-decision-engine has a growth score of 4.44 and 44 stars, aiming for an open-source reproduction of TypeSafe Jev with improved calibration and speed at a fraction of the computational cost compared to its proprietary counterpart. The project's growth reflects interest in accessible alternatives that offer performance enhancements over existing solutions.
Lastly, Thinking-Space/One-Shot-OPD has a growth score of 2.48 and 85 stars, rethinking on-policy distillation techniques for large language models with an emphasis on one-shot training examples. The project's steady growth highlights the ongoing exploration in efficient training methods that can significantly reduce resource requirements while maintaining model performance.
These projects collectively showcase the evolving landscape of AI fine-tuning and training, from decision-making engines to dataset curation tools and novel approaches to LLM training, underscoring the dynamic nature of current research and development efforts in this space.
Mapika/decider stands out this week with a growth score of 61.54 and 838 stars. It provides one-pass typed decisions with calibrated probabilities inspired by System One style models, fine-tuned from Qwen3.5-2B. The project's popularity likely stems from its innovative approach to decision-making in AI systems, offering faster and more reliable outcomes compared to conventional methods.
Liuziyu77/Valen is another prominent tool with a growth score of 48.60 and 336 stars. This repository aims to train a Jev-like multimodal model that integrates vision capabilities into decision-making processes, expanding the scope of TypeSafe Jev models. The project's rapid growth can be attributed to its unique approach to integrating visual data with AI decision engines, making it highly relevant for applications requiring both textual and image-based inputs.
AkashPriyadarshii/jev-curate has a growth score of 26.95 and 86 stars. This tool is designed to sift through high-throughput synthetic and pretraining datasets for TypeSafe Jev models, using Rust streaming core technology with Parquet and JSONL I/O formats. The project's growing popularity can be linked to its efficient dataset processing capabilities that significantly enhance the training efficiency of Jev models.
davefano/omarchy-trackpad-plus has a growth score of 13.31 and 126 stars, focusing on fine-grained trackpad controls for Omarchy devices. While not directly related to AI model training, it shows promise in enhancing user interaction with AI-driven systems through precise device customization.
friday-memory/friday, with a growth score of 11.46 and 62 stars, introduces an open-source persistent cognitive memory layer designed to prevent "amnesia" in AI coding agents like Cursor, Claude & Copilot. This project's rising star count suggests growing interest in maintaining context awareness in AI-driven development environments.
luigisaetta/llm-fine-tuning-on-mac has a growth score of 5.56 and 21 stars, offering fine-tuning capabilities for small LLMs using LoRA on Mac devices. The project's modest but steady growth indicates its relevance to developers looking to perform model tuning locally without high-end hardware requirements.
intikhab49/open-jev-typed-decision-engine has a growth score of 4.44 and 44 stars, aiming for an open-source reproduction of TypeSafe Jev with improved calibration and speed at a fraction of the computational cost compared to its proprietary counterpart. The project's growth reflects interest in accessible alternatives that offer performance enhancements over existing solutions.
Lastly, Thinking-Space/One-Shot-OPD has a growth score of 2.48 and 85 stars, rethinking on-policy distillation techniques for large language models with an emphasis on one-shot training examples. The project's steady growth highlights the ongoing exploration in efficient training methods that can significantly reduce resource requirements while maintaining model performance.
These projects collectively showcase the evolving landscape of AI fine-tuning and training, from decision-making engines to dataset curation tools and novel approaches to LLM training, underscoring the dynamic nature of current research and development efforts in this space.