Today's Fine-tuning & Training: Fastest-Growing Projects — September 26, 2026
Today's the Fine-tuning & Training space, we've seen a continued focus on multimodal models and decision engines that leverage non-autoregressive passes for efficiency gains. The Mapika/decider project leads with a strong growth score, indicating robust community interest in its approach to one-pass typed decisions.
Mapika/decider is a model fine-tuned from Qwen3.5-2B designed for making decisions based on calibrated probabilities using a System One style model. Its high Growth Score of 50.30 and 424 stars suggest that developers are keenly interested in its efficiency and performance improvements over traditional decision-making models.
Liuziyu77/Valen offers the ability to train Jev-like multimodal models with vision capabilities, extending beyond text-only processing. With a Growth Score of 43.67 and 163 stars, it's clear that this project is gaining traction as researchers and developers seek more integrated visual and textual processing solutions.
AkashPriyadarshii/jev-curate provides high-throughput synthetic and pretraining dataset filtering for TypeSafe Jev, with Rust streaming core capabilities. Its Growth Score of 32.69 and 73 stars indicate its utility in accelerating the training process through efficient data handling and high throughput rates.
davefano/omarchy-trackpad-plus focuses on fine-grained trackpad controls and pointer-feel tuning for Omarchy, which is somewhat tangential to mainstream AI fine-tuning but still garners interest with 125 stars. However, its Growth Score of 17.11 suggests steady development rather than explosive growth.
friday-memory/friday introduces a persistent cognitive memory layer designed to enhance the functionality of AI coding agents like Claude and Copilot by ensuring they retain context across sessions. Its Growth Score of 13.60 and 59 stars reflect growing interest in its potential for improving the continuity and effectiveness of AI interactions in development environments.
luigisaetta/llm-fine-tuning-on-mac provides all necessary code to fine-tune a small language model using LoRA on macOS, aiming at making LLMs more accessible to developers. Despite having fewer stars (21) compared to others, its Growth Score of 6.00 indicates continuous development and some interest in practical tools for fine-tuning models locally.
intikhab49/open-jev-typed-decision-engine offers an open-source reproduction of a TypeSafe Jev model with improved calibration and speed over the original Jev. This project, with a Growth Score of 5.71 and 44 stars, highlights efforts to democratize access to high-performance decision engines through efficient training methods.
Thinking-Space/One-Shot-OPD explores on-policy distillation techniques for large language models using only one training example per iteration. With a modest Growth Score of 2.67 but still attracting 84 stars, this project is intriguing for its potential in optimizing the fine-tuning process and reducing data requirements.
These projects collectively demonstrate an active community focused on advancing decision-making capabilities, multimodal processing, efficient dataset handling, and novel training methodologies within the AI ecosystem.
Mapika/decider is a model fine-tuned from Qwen3.5-2B designed for making decisions based on calibrated probabilities using a System One style model. Its high Growth Score of 50.30 and 424 stars suggest that developers are keenly interested in its efficiency and performance improvements over traditional decision-making models.
Liuziyu77/Valen offers the ability to train Jev-like multimodal models with vision capabilities, extending beyond text-only processing. With a Growth Score of 43.67 and 163 stars, it's clear that this project is gaining traction as researchers and developers seek more integrated visual and textual processing solutions.
AkashPriyadarshii/jev-curate provides high-throughput synthetic and pretraining dataset filtering for TypeSafe Jev, with Rust streaming core capabilities. Its Growth Score of 32.69 and 73 stars indicate its utility in accelerating the training process through efficient data handling and high throughput rates.
davefano/omarchy-trackpad-plus focuses on fine-grained trackpad controls and pointer-feel tuning for Omarchy, which is somewhat tangential to mainstream AI fine-tuning but still garners interest with 125 stars. However, its Growth Score of 17.11 suggests steady development rather than explosive growth.
friday-memory/friday introduces a persistent cognitive memory layer designed to enhance the functionality of AI coding agents like Claude and Copilot by ensuring they retain context across sessions. Its Growth Score of 13.60 and 59 stars reflect growing interest in its potential for improving the continuity and effectiveness of AI interactions in development environments.
luigisaetta/llm-fine-tuning-on-mac provides all necessary code to fine-tune a small language model using LoRA on macOS, aiming at making LLMs more accessible to developers. Despite having fewer stars (21) compared to others, its Growth Score of 6.00 indicates continuous development and some interest in practical tools for fine-tuning models locally.
intikhab49/open-jev-typed-decision-engine offers an open-source reproduction of a TypeSafe Jev model with improved calibration and speed over the original Jev. This project, with a Growth Score of 5.71 and 44 stars, highlights efforts to democratize access to high-performance decision engines through efficient training methods.
Thinking-Space/One-Shot-OPD explores on-policy distillation techniques for large language models using only one training example per iteration. With a modest Growth Score of 2.67 but still attracting 84 stars, this project is intriguing for its potential in optimizing the fine-tuning process and reducing data requirements.
These projects collectively demonstrate an active community focused on advancing decision-making capabilities, multimodal processing, efficient dataset handling, and novel training methodologies within the AI ecosystem.