Today's Fine-tuning & Training: Fastest-Growing Projects — September 25, 2026
Today's the Fine-tuning & Training space on GitHub, there's a noticeable trend towards more efficient and versatile model training techniques, with projects focusing on multi-modal capabilities, high-throughput data processing, and fine-grained control over AI decision-making processes. The Mapika/decider project leads this category with its unique approach to one-pass typed decisions using calibrated probabilities, demonstrating significant growth in the past month.
Mapika/decider is a System One style model that fine-tunes from Qwen3.5-2B for making swift and accurate decisions based on probabilistic assessments. Its substantial growth score of 51.00 and rising star count to 354 indicate active community engagement and interest in its innovative approach to decision-making models.
Liuziyu77/Valen aims to train a Jev-like multimodal model that includes vision capabilities, allowing for more comprehensive and context-aware decisions. With a growth score of 35.00 and 83 stars, Valen's traction suggests it is gaining popularity among developers looking to enhance the visual understanding aspect of AI decision-making.
AkashPriyadarshii/jev-curate focuses on high-throughput synthetic and pretraining dataset processing for TypeSafe Jev models, featuring a Rust streaming core and Parquet and JSONL I/O capabilities. Despite its lower growth score of 25.57, the project's 44 commits in the last month and 64 stars highlight ongoing development efforts and community interest in efficient data processing solutions.
davefano/omarchy-trackpad-plus is designed for fine-grained trackpad control tuning within Omarchy, though its focus on hardware rather than software training might set it apart from other projects. The project's growth score of 18.27 and 124 stars indicate sustained interest among developers looking to refine their device interactions.
friday-memory/friday introduces a persistent cognitive memory layer for AI coding agents like Claude & Copilot, helping them maintain context across sessions. With a growth score of 15.06 and 58 stars, the project's steady development pace (32 commits in the last month) reflects ongoing improvements to enhance AI agent performance.
intikhab49/open-jev-typed-decision-engine offers an open-source reproduction of TypeSafe Jev with a focus on efficient decision-making through non-autoregressive passes. The project’s 6.58 growth score and 43 stars suggest it is attracting attention from developers interested in exploring more calibrated and faster decision engines.
luigisaetta/llm-fine-tuning-on-mac provides all the necessary code for fine-tuning a small LLM with LoRA on macOS, catering to developers seeking localized model optimization. With a growth score of 6.25 and 21 stars, this project's recent spike in activity (45 commits) highlights increased interest in efficient local training solutions.
Thinking-Space/One-Shot-OPD explores an innovative approach to on-policy distillation for large language models using only one training example, aiming to optimize model efficiency. Despite a lower growth score of 2.77 and 83 stars, the project's steady development (5 commits in the last month) suggests ongoing research interest in minimalistic training techniques.
These projects collectively underscore the diversity and innovation within the Fine-tuning & Training space, each addressing unique challenges with novel solutions that are gaining traction among developers and researchers alike.
Mapika/decider is a System One style model that fine-tunes from Qwen3.5-2B for making swift and accurate decisions based on probabilistic assessments. Its substantial growth score of 51.00 and rising star count to 354 indicate active community engagement and interest in its innovative approach to decision-making models.
Liuziyu77/Valen aims to train a Jev-like multimodal model that includes vision capabilities, allowing for more comprehensive and context-aware decisions. With a growth score of 35.00 and 83 stars, Valen's traction suggests it is gaining popularity among developers looking to enhance the visual understanding aspect of AI decision-making.
AkashPriyadarshii/jev-curate focuses on high-throughput synthetic and pretraining dataset processing for TypeSafe Jev models, featuring a Rust streaming core and Parquet and JSONL I/O capabilities. Despite its lower growth score of 25.57, the project's 44 commits in the last month and 64 stars highlight ongoing development efforts and community interest in efficient data processing solutions.
davefano/omarchy-trackpad-plus is designed for fine-grained trackpad control tuning within Omarchy, though its focus on hardware rather than software training might set it apart from other projects. The project's growth score of 18.27 and 124 stars indicate sustained interest among developers looking to refine their device interactions.
friday-memory/friday introduces a persistent cognitive memory layer for AI coding agents like Claude & Copilot, helping them maintain context across sessions. With a growth score of 15.06 and 58 stars, the project's steady development pace (32 commits in the last month) reflects ongoing improvements to enhance AI agent performance.
intikhab49/open-jev-typed-decision-engine offers an open-source reproduction of TypeSafe Jev with a focus on efficient decision-making through non-autoregressive passes. The project’s 6.58 growth score and 43 stars suggest it is attracting attention from developers interested in exploring more calibrated and faster decision engines.
luigisaetta/llm-fine-tuning-on-mac provides all the necessary code for fine-tuning a small LLM with LoRA on macOS, catering to developers seeking localized model optimization. With a growth score of 6.25 and 21 stars, this project's recent spike in activity (45 commits) highlights increased interest in efficient local training solutions.
Thinking-Space/One-Shot-OPD explores an innovative approach to on-policy distillation for large language models using only one training example, aiming to optimize model efficiency. Despite a lower growth score of 2.77 and 83 stars, the project's steady development (5 commits in the last month) suggests ongoing research interest in minimalistic training techniques.
These projects collectively underscore the diversity and innovation within the Fine-tuning & Training space, each addressing unique challenges with novel solutions that are gaining traction among developers and researchers alike.