Trelis Research

Trelis offers done-for-you custom model fine-tuning services, with a focus on voice models (TTS, STT).

• Access to all SEVEN Trelis Repos (ADVANCED- Voice, Vision, Fine-tuning (LLMs), Inference, Evals, Time-Series, and Robotics)
• Support via Github Issues and Trelis’ Private Discord.

• Dataset filtering and preparation techniques.
• Synthetic data generation methods.
• Fine-tuning methods including LoRA, full fine-tuning, DPO, ORPO.
• Support for Open Source (Llama, Qwen etc.) and OpenAI models.
• Single and multi-GPU training (including DDP and FSDP).

• Dataset preparation for transcription, voice cloning or multi-modal audio + text models
• Fine-tune multi-modal Audio + Text Models (Qwen 2 Audio)
• Speech-to-Text Transcription (Fine-tuning and Serving Whisper Turbo)
• Text-to-Speech / Voice Cloning (Fine-tuning StyleTTS2)
• Run Speech-to-Speech models locally or in the cloud

• Data Preparation and Training Guides
• ACT (Action Chunking Transformer)
• GR00T-N1

• Dataset Preparation for multi-modal image + text models OR diffusion models.
• Fine-tuning diffusion models for image generation with FLUX Schnell or FLUX dev.
• Fine-tuning vision models for custom image OR text datasets. [Qwen VL, Pixtral, Florence 2, LLaVA].
• Fine-tune vision models for bounding box detection.
• Deploying a server / api endpoint for Qwen VL.

• Creating evaluation datasets
• Setting up LLM as a judge
• Measuring the performance of your AI system (i.e. running “evals”)

• Forecasting with transformers (e.g. weather, power demand, prices)
• Forecasting and Evaluation Scripts
• Training and Fine-tuning Scripts

• API Setup guides: RunPod, VastAI or your own laptop | SGLang, Nvidia NIM, vLLM, TGI, llama.cpp
• Serverless Inference Guide.
• Inference speed-up techniques: Context caching, speculative decoding, quantization, output predictions.
• Function calling and structured data extraction (e.g. json) methods.
• Sensitive data redaction (with presidio).
• Security precautions (against jailbreaking and prompt injection).
• Build an AI assistant (RAG pipeline) including vector search, bm25 and citation verification.

• YouTube Videos: Post a comment on YouTube. I respond to most comments (same for comments on Substack or X).
• Questions about Github Repos/Scripts: Post a question on the corresponding product page (“Learn More” links above).
• Github Repo Support (after purchase): Create an issue in the relevant GitHub repo.
• Other topics/questions: You can purchase lifetime access to the Trelis Discord here. Access is also included when you buy the Github Repo Bundle.