Today's Fine-tuning & Training: Fastest-Growing Projects — September 27, 2026
Today's the Fine-tuning & Training space, there's a noticeable trend towards systems that streamline decision-making processes with calibrated probabilities and those designed to handle multimodal training efficiently. One standout project focusing on single-pass decisions is Mapika/decider, which has seen significant growth due to its unique approach to model fine-tuning from Qwen3.5-2B. With over 500 stars, it exemplifies the demand for decision-making models that can operate with high efficiency and accuracy.
Mapika/decider: This project offers one-pass typed decisions with calibrated probabilities, fine-tuned from Qwen3.5-2B. Its impressive growth score of 51.05 and a substantial number of stars (523) indicate strong community interest in its innovative approach to decision-making models.
Liuziyu77/Valen: Train a Jev-like multimodal model by yourself, extending the System One Model with vision capabilities. The project's growth score of 48.12 highlights its relevance among developers looking for advanced multimodal training solutions that integrate visual data effectively.
AkashPriyadarshii/jev-curate: This tool is a high-throughput synthetic and pretraining dataset sifter designed for TypeSafe Jev, with Rust streaming core capabilities and support for Parquet and JSONL I/O formats. Its steady growth score of 29.72, along with 70 commits in the last month, suggests active development and interest from users seeking efficient data processing solutions.
davefano/omarchy-trackpad-plus: This project focuses on fine-grained trackpad controls for Omarchy, offering precision tuning options for devices. Although its focus is more specific to hardware customization than general AI training, it still garners significant attention with 126 stars and a growth score of 14.40.
friday-memory/friday: This project introduces an open-source persistent cognitive memory layer designed to enhance the performance of AI coding agents like Claude & Copilot by ensuring they retain context across sessions. Its growth score of 12.41, coupled with 60 stars, indicates a growing need for such memory solutions in the development environment.
luigisaetta/llm-fine-tuning-on-mac: This repository provides detailed code and instructions for fine-tuning small LLMs using LoRA on macOS systems. The project's modest growth score of 5.77 reflects its niche appeal to developers interested in local machine learning model tuning on their personal devices.
intikhab49/open-jev-typed-decision-engine: Aiming to reproduce the TypeSafe Jev decision engine, this project offers a more efficient and faster alternative with better-calibrated confidence scores compared to the original. Its growth score of 5.00 and 44 stars suggest ongoing interest from developers looking for open-source alternatives in decision-making models.
Thinking-Space/One-Shot-OPD: This repository explores on-policy distillation techniques for large language models, focusing on training with just one example per model. With a growth score of 2.58 and 85 stars, it demonstrates a growing interest in optimizing the training process through innovative algorithmic approaches.
Today's projects highlight diverse efforts within the fine-tuning & training domain, ranging from specialized hardware tuning to advanced multimodal models and efficient data processing systems for decision-making engines.
Mapika/decider: This project offers one-pass typed decisions with calibrated probabilities, fine-tuned from Qwen3.5-2B. Its impressive growth score of 51.05 and a substantial number of stars (523) indicate strong community interest in its innovative approach to decision-making models.
Liuziyu77/Valen: Train a Jev-like multimodal model by yourself, extending the System One Model with vision capabilities. The project's growth score of 48.12 highlights its relevance among developers looking for advanced multimodal training solutions that integrate visual data effectively.
AkashPriyadarshii/jev-curate: This tool is a high-throughput synthetic and pretraining dataset sifter designed for TypeSafe Jev, with Rust streaming core capabilities and support for Parquet and JSONL I/O formats. Its steady growth score of 29.72, along with 70 commits in the last month, suggests active development and interest from users seeking efficient data processing solutions.
davefano/omarchy-trackpad-plus: This project focuses on fine-grained trackpad controls for Omarchy, offering precision tuning options for devices. Although its focus is more specific to hardware customization than general AI training, it still garners significant attention with 126 stars and a growth score of 14.40.
friday-memory/friday: This project introduces an open-source persistent cognitive memory layer designed to enhance the performance of AI coding agents like Claude & Copilot by ensuring they retain context across sessions. Its growth score of 12.41, coupled with 60 stars, indicates a growing need for such memory solutions in the development environment.
luigisaetta/llm-fine-tuning-on-mac: This repository provides detailed code and instructions for fine-tuning small LLMs using LoRA on macOS systems. The project's modest growth score of 5.77 reflects its niche appeal to developers interested in local machine learning model tuning on their personal devices.
intikhab49/open-jev-typed-decision-engine: Aiming to reproduce the TypeSafe Jev decision engine, this project offers a more efficient and faster alternative with better-calibrated confidence scores compared to the original. Its growth score of 5.00 and 44 stars suggest ongoing interest from developers looking for open-source alternatives in decision-making models.
Thinking-Space/One-Shot-OPD: This repository explores on-policy distillation techniques for large language models, focusing on training with just one example per model. With a growth score of 2.58 and 85 stars, it demonstrates a growing interest in optimizing the training process through innovative algorithmic approaches.
Today's projects highlight diverse efforts within the fine-tuning & training domain, ranging from specialized hardware tuning to advanced multimodal models and efficient data processing systems for decision-making engines.