🚀 LLM Optimization & Specialized Training Hub

A centralized repository of high-performance fine-tuning recipes, reasoning models, and edge deployment optimizations developed for real-world production environments.

📂 Project Categories

Reasoning & RLHF

Implementing Group Relative Policy Optimization (GRPO) for logic and math reasoning without a separate Critic model.

• Reward functions: Logic and math reasoning without a separate Critic model, popularized by DeepSeek-R1....
• Qwen3 8B FP8 GRPO: Logic and math reasoning without a separate Critic model, popularized by DeepSeek-R1....
• Reward functions: Logic and math reasoning without a separate Critic model, popularized by DeepSeek-R1....
• Qwen3 (14B)-Reasoning-Conversational: High-performance LLM optimization implemented with Unsloth kernels....

Efficient Architectures

Optimizing non-dense models (Mixture-of-Experts) and modern encoders (ModernBERT) for high throughput and classification.

• TinyQwen3 MoE: To demonstrate unsloth's MoE optimisations on T4 for imdatta0/tiny_qwen3_moe_2.8B_0.7B by finetuni...
• ModernBert: High-performance LLM optimization implemented with Unsloth kernels....
• Qwen3 MoE: To convert unsloth/Qwen3-30B-A3B-Instruct-2507 into a reasoning model via GRPO by using OpenR1's M...
• bert classification: High-performance LLM optimization implemented with Unsloth kernels....

Edge & Quantization

Shrinking models to run on phones and low-power devices using Quantization Aware Training (QAT) and sub-1B parameter scaling.

• Qwen3 (0 6B)-Phone Deployment: High-performance LLM optimization implemented with Unsloth kernels....
• Qwen3 (4B) Instruct-QAT: High-performance LLM optimization implemented with Unsloth kernels....
• Nemotron-3-Nano-30B-A3B A100: High-performance LLM optimization implemented with Unsloth kernels....
• Phi 4-Conversational: High-performance LLM optimization implemented with Unsloth kernels....

Multimodal & Audio

Beyond text: Fine-tuning robust speech-to-text and high-fidelity text-to-speech synthesis models.

• Whisper: Optimized multimodal processing for low-latency audio/speech tasks....
• Oute TTS (1B): Optimized multimodal processing for low-latency audio/speech tasks....
• Orpheus (3B)-TTS: Optimized multimodal processing for low-latency audio/speech tasks....
• Spark TTS (0 5B): Optimized multimodal processing for low-latency audio/speech tasks....

Synthetic Data

Data curation and distillation. Using large models to generate high-quality instruction sets to improve smaller specialized models.

• Synthetic Data Hackathon: High-performance LLM optimization implemented with Unsloth kernels....
• Meta Synthetic Data Llama3 2 (3B): High-performance LLM optimization implemented with Unsloth kernels....

🛤️ Moving to Production

These implementations are designed for 2026-scale workloads. To deploy:

1. Distillation: Use the Synthetic Data pipelines to transfer logic from Llama 3.2 to smaller models.
2. Quantization: Apply QAT or FP8 post-training quantization for edge suitability.
3. Deployment: Export to GGUF or vLLM for high-throughput serving.