lora

Tuning the Finetuning: An exploration of achieving success with QLoRA

Privacy first powerful, browser-based tool for tagging, captioning, cropping and managing training datasets for Stable Diffusion's LoRA trainings. No install required.

Local-first desktop manager for AI image libraries

telegram bot for self-hosted local inference of stable diffusion, text-to-speech and large language models, such as llama3

Easy wrapper for inserting LoRA layers in CLIP.

Training LLaMA language model with MMEngine! It supports LoRA fine-tuning!

Preprint: Asymmetry in Low-Rank Adapters of Foundation Models

Custom model training using modern architectures. 4-bit QLoRA fine-tuning pipeline for LLaMA 3 8B with production-grade optimization. Memory-efficient training on consumer GPUs. Published adapter on HuggingFace. From training pipeline to deployed model.

Data Curation Tool is a local-first dataset curation, tagging, model-management, downloader, 3D generation, media metadata, and tool-orchestration application. The current documentation set is stored in docs/wiki/

SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models

Synthetic medical VQA pipeline: 119K images annotated by frontier VLMs, cross-validated at 93% agreement, fine-tuned on 3 model families (2-3B params)

Train YOLO + VLM with one command. Auto-generate vision-language training data from YOLO labels - no extra labeling needed.

在 ComfyUI 里获得 WebUI 式提示词体验：中文翻译、Tag 自动补全、LoRA 管理、提示词市场和高级模块接入。

One GPU. Full LLM workflow. Real benchmarks. No cloud required.

Prompt Engineering for Language models (GPT-3, GPT-4, chatGPT) and text-to-image models (Stable Diffusion, Midjourney, Dall-e)

Train LoRA using Microsoft's official implementation with Stable Diffusion models.

Fine-tuned Qwen2-VL-7B for LaTeX OCR using LoRA and Unsloth on the LaTeX OCR dataset. Built augmentation pipeline (rotation, noise, contrast jitter), ran LoRA rank sweep (r=8/16/32), and evaluated across CER, Token F1, BLEU-4, and Exact Match. Deployed as a Gradio Space with live metric computation.

🔥 (SIGGRAPH 2026) 3D-aware hairstyle transfer for images: in-context bald conversion, pose-aligned hair import, and identity-preserving synthesis.

LoRA Manager for Civit Models

PipelineLLM 是一个系统性的大语言模型（LLM）后训练学习项目，涵盖从监督微调（SFT）到偏好优化（DPO）、强化学习（RLHF/PPO/GRPO）再到持续学习（Continual Learning)的完整技术栈。