Today's Fine-tuning & Training: Fastest-Growing Projects — June 14, 2026
Today's trend in the Fine-tuning & Training space highlights a growing interest in specialized solutions for large language models (LLMs) and speech processing, alongside practical frameworks that streamline model fine-tuning on local hardware. One of the standout projects is "speechkv-trim," which offers innovative pruning techniques tailored to long-form speech LLMs.
jelllott/speechkv-trim: This repository introduces Speech-aware KV cache pruning for optimizing long-form speech LLMs like Qwen2-Audio and SALMONN, focusing on token/head/chunk-level pruners evaluated against LibriSpeech-long and GigaSpeech datasets. With a strong Growth Score of 33.29 and 219 stars, it indicates significant interest in enhancing the efficiency and performance of speech-based LLMs.
pat-jj/harness-1: This project provides an ultra recipe for training long-horizon search agents that can match frontier AI's search capabilities using a 20B model. The repository has garnered substantial attention with over 597 stars, reflecting its relevance in advancing the state-of-the-art in large-scale agent training and optimization.
DaoyuanLi2816/can-i-finetune-this: This tool estimates whether Hugging Face models can fit and be fine-tuned on local GPUs, aiding researchers and developers in resource management. With a Growth Score of 13.95 and 467 stars, it shows promise in addressing practical challenges faced by users looking to work with large models on constrained hardware.
zengxiao-he/tessera: From teacher to tiles — this project is an LLM distillation and serving engine that leverages custom Triton/CUDA kernels and FSDP distillation techniques. Its Growth Score of 12.78 and 155 stars suggest it's gaining traction for its comprehensive approach in model optimization and deployment.
Mengqi-Lei/count-anything: This repository includes code and implementation guidelines based on the paper "Counting Anything," providing detailed instructions and resources for researchers aiming to apply counting tasks with AI models. With a Growth Score of 4.76, it continues to attract interest from those working on object detection and semantic segmentation projects.
gvkhosla/pi-tinker: This project enables users to fine-tune open-source models within Pi using Tinker, offering managed improve loops, data preparation tools, evaluations, smoke tests, and deployment snippets. Its Growth Score of 4.23 and 21 stars indicate steady growth for its user-friendly approach to model refinement.
thombanal/clip-finetune-recipes: Offering practical CLIP fine-tuning recipes, this repository covers distributed training, LoRA techniques, hard-negative mining, and leakage checks. Despite having no recent commits, it remains popular with 221 stars, reflecting the enduring interest in optimizing image-text matching tasks.
SoloCalm/MiniLoRA: This project provides a Qwen2.5-0.5B medical LLM fine-tuning tutorial using LoRA techniques. Its Growth Score of 2.63 and modest 31 stars suggest it's gaining traction among those interested in applying fine-tuning methods to specific domains such as healthcare.
JuliusBrussee/cavegemma: This repository aims to fine-tune the Gemma 4 31B model to produce caveman-style speech natively, showcasing creative approaches to language generation. With a Growth Score of 1.79 and 47 stars, it highlights the community's interest in experimenting with stylistic variations in LLM outputs.
These tools collectively showcase the diversity and depth of innovation in the fine-tuning and training domain, from specialized speech optimizations to broad-reaching agent training methodologies.
jelllott/speechkv-trim: This repository introduces Speech-aware KV cache pruning for optimizing long-form speech LLMs like Qwen2-Audio and SALMONN, focusing on token/head/chunk-level pruners evaluated against LibriSpeech-long and GigaSpeech datasets. With a strong Growth Score of 33.29 and 219 stars, it indicates significant interest in enhancing the efficiency and performance of speech-based LLMs.
pat-jj/harness-1: This project provides an ultra recipe for training long-horizon search agents that can match frontier AI's search capabilities using a 20B model. The repository has garnered substantial attention with over 597 stars, reflecting its relevance in advancing the state-of-the-art in large-scale agent training and optimization.
DaoyuanLi2816/can-i-finetune-this: This tool estimates whether Hugging Face models can fit and be fine-tuned on local GPUs, aiding researchers and developers in resource management. With a Growth Score of 13.95 and 467 stars, it shows promise in addressing practical challenges faced by users looking to work with large models on constrained hardware.
zengxiao-he/tessera: From teacher to tiles — this project is an LLM distillation and serving engine that leverages custom Triton/CUDA kernels and FSDP distillation techniques. Its Growth Score of 12.78 and 155 stars suggest it's gaining traction for its comprehensive approach in model optimization and deployment.
Mengqi-Lei/count-anything: This repository includes code and implementation guidelines based on the paper "Counting Anything," providing detailed instructions and resources for researchers aiming to apply counting tasks with AI models. With a Growth Score of 4.76, it continues to attract interest from those working on object detection and semantic segmentation projects.
gvkhosla/pi-tinker: This project enables users to fine-tune open-source models within Pi using Tinker, offering managed improve loops, data preparation tools, evaluations, smoke tests, and deployment snippets. Its Growth Score of 4.23 and 21 stars indicate steady growth for its user-friendly approach to model refinement.
thombanal/clip-finetune-recipes: Offering practical CLIP fine-tuning recipes, this repository covers distributed training, LoRA techniques, hard-negative mining, and leakage checks. Despite having no recent commits, it remains popular with 221 stars, reflecting the enduring interest in optimizing image-text matching tasks.
SoloCalm/MiniLoRA: This project provides a Qwen2.5-0.5B medical LLM fine-tuning tutorial using LoRA techniques. Its Growth Score of 2.63 and modest 31 stars suggest it's gaining traction among those interested in applying fine-tuning methods to specific domains such as healthcare.
JuliusBrussee/cavegemma: This repository aims to fine-tune the Gemma 4 31B model to produce caveman-style speech natively, showcasing creative approaches to language generation. With a Growth Score of 1.79 and 47 stars, it highlights the community's interest in experimenting with stylistic variations in LLM outputs.
These tools collectively showcase the diversity and depth of innovation in the fine-tuning and training domain, from specialized speech optimizations to broad-reaching agent training methodologies.