Today's Fine-tuning & Training: Fastest-Growing Projects — June 15, 2026
Today's the Fine-tuning & Training category, there's a noticeable trend towards optimizing large language models (LLMs) through various methods such as distillation, quantization, and cache pruning, highlighting ongoing efforts to enhance both efficiency and performance. The top tool this week is "harness-1" by pat-jj, which has seen significant growth in popularity.
pat-jj/harness-1
This repository provides an ultra recipe for training long-horizon search agents with a 20B model, aiming to match the advanced search capabilities of frontier AI systems. With its impressive Growth Score of 17.29 and over 600 stars, it stands out as a highly active project focusing on enhancing the scalability and performance of LLMs.
zengxiao-he/tessera
Tessera is a comprehensive solution for distilling large language models into smaller ones through custom Triton/CUDA kernels and FSDP distillation techniques. The growth in this repository, with a Growth Score of 12.10 and 167 stars, reflects its utility in serving efficiently distilled LLMs.
ZunhaiSu/OScaR-KV-Quant
OScaR introduces innovative KV cache quantization for large language models to achieve better accuracy-efficiency trade-offs. With a Growth Score of 8.09 and 134 stars, this project is gaining traction due to its unique approach in optimizing model performance without sacrificing accuracy.
Mengqi-Lei/count-anything
This repository offers code and guidelines for counting objects accurately using machine learning models, as described in the associated paper "Counting Anything." With a Growth Score of 4.61 and 108 stars, it demonstrates ongoing interest in applying advanced techniques to specific tasks like object counting.
gvkhosla/pi-tinker
Pi-Tinker enables users to fine-tune open-source models inside Pi with managed improve loops, data preparation, and evaluation tools, simplifying the process of deploying and evaluating model improvements. Its Growth Score of 3.88 and modest 21 stars indicate steady interest from developers looking for streamlined fine-tuning solutions.
jelllott/speechkv-trim
Focused on speech-aware KV cache pruning for long-form speech LLMs, this project aims to optimize these models by implementing token-level pruners evaluated against standard datasets. With a Growth Score of 3.37 and 219 stars, it shows promise in addressing the unique challenges posed by speech data.
SoloCalm/MiniLoRA
This repository provides tutorials for fine-tuning Qwen2.5-0.5B models specifically for medical applications using LoRA techniques. Its Growth Score of 2.54 and 31 stars suggest that it is a valuable resource for those interested in applying LLMs to healthcare contexts.
These projects collectively highlight the diverse approaches developers are taking to optimize and fine-tune large language models, from distillation and quantization to specialized applications like speech recognition and medical diagnostics.
pat-jj/harness-1
This repository provides an ultra recipe for training long-horizon search agents with a 20B model, aiming to match the advanced search capabilities of frontier AI systems. With its impressive Growth Score of 17.29 and over 600 stars, it stands out as a highly active project focusing on enhancing the scalability and performance of LLMs.
zengxiao-he/tessera
Tessera is a comprehensive solution for distilling large language models into smaller ones through custom Triton/CUDA kernels and FSDP distillation techniques. The growth in this repository, with a Growth Score of 12.10 and 167 stars, reflects its utility in serving efficiently distilled LLMs.
ZunhaiSu/OScaR-KV-Quant
OScaR introduces innovative KV cache quantization for large language models to achieve better accuracy-efficiency trade-offs. With a Growth Score of 8.09 and 134 stars, this project is gaining traction due to its unique approach in optimizing model performance without sacrificing accuracy.
Mengqi-Lei/count-anything
This repository offers code and guidelines for counting objects accurately using machine learning models, as described in the associated paper "Counting Anything." With a Growth Score of 4.61 and 108 stars, it demonstrates ongoing interest in applying advanced techniques to specific tasks like object counting.
gvkhosla/pi-tinker
Pi-Tinker enables users to fine-tune open-source models inside Pi with managed improve loops, data preparation, and evaluation tools, simplifying the process of deploying and evaluating model improvements. Its Growth Score of 3.88 and modest 21 stars indicate steady interest from developers looking for streamlined fine-tuning solutions.
jelllott/speechkv-trim
Focused on speech-aware KV cache pruning for long-form speech LLMs, this project aims to optimize these models by implementing token-level pruners evaluated against standard datasets. With a Growth Score of 3.37 and 219 stars, it shows promise in addressing the unique challenges posed by speech data.
SoloCalm/MiniLoRA
This repository provides tutorials for fine-tuning Qwen2.5-0.5B models specifically for medical applications using LoRA techniques. Its Growth Score of 2.54 and 31 stars suggest that it is a valuable resource for those interested in applying LLMs to healthcare contexts.
These projects collectively highlight the diverse approaches developers are taking to optimize and fine-tune large language models, from distillation and quantization to specialized applications like speech recognition and medical diagnostics.