Roadmap | LLM Dojo

00

Foundations & Environment

Set up your environment, understand transformers from scratch, and master data preparation fundamentals.

00Environment Setup 01Transformer Basics 02Tokenization Deep Dive 03Dataset Preparation Basics 04Data Quality Analysis 05Handling Class Imbalance 06Data Augmentation for NLP

01

Full Model Fine-Tuning

Fine-tune complete transformer models for classification and generation. Build custom training loops with mixed precision.

10 notebooks →

10Full Fine-Tuning: Classification 11Full Fine-Tuning: Generation 12Custom Training Loop 13Custom Loss Functions 14Imbalanced Classification 15Mixed Precision Training 16Gradient Accumulation 17Regularization Techniques 18Evaluation Metrics 19Curriculum Learning

02

Parameter-Efficient Fine-Tuning

Master LoRA, QLoRA, adapters, and prompt tuning — train LLMs with 100x fewer parameters.

10 notebooks →

20LoRA Theory 21LoRA on LLaMA-2 7B 22QLoRA Implementation 23LoRA Target Modules 24Custom Loss with LoRA 25Adapter Layers 26Prompt Tuning 27PEFT Method Comparison 28LoRA Merging 29Advanced LoRA Variants

03

Advanced Optimization

FlashAttention, DeepSpeed ZeRO, FSDP, gradient checkpointing, and instruction tuning at scale.

10 notebooks →

30Flash Attention 31DeepSpeed ZeRO 32FSDP PyTorch 33Gradient Checkpointing 34Optimizer Comparison 35LR Schedules 36Advanced Data Loading 37Instruction Tuning 38Long Context Training 39Contrastive Learning

04

Alignment & Specialized Techniques

RLHF, DPO, Constitutional AI, reward models, and safety evaluation for aligned LLMs.

10 notebooks →

40Reward Model Training 41RLHF with PPO 42DPO Implementation 43Constitutional AI 44Domain Adaptation 45Multi-Task Fine-Tuning 46Catastrophic Forgetting 47Negative Sampling Strategies 48Safety Evaluation 49Mixture of Experts

05

Custom Kernels & Production

Write CUDA/Triton kernels, master quantization, implement speculative decoding, and deploy with vLLM.

10 notebooks →

50CUDA Basics 51Triton Kernels 52Custom Attention Kernel 53Fused Operations 54Quantization Methods 55KV Cache Optimization 56Speculative Decoding 57vLLM Serving 58TensorRT-LLM 59Continuous Batching

06

LLM Inference Optimization

Profile, optimize, and deploy LLM inference at scale — from KV cache to quantization to multi-GPU serving.

20 notebooks →

07

2024–2025 Techniques

GRPO reasoning models, ORPO/KTO alignment, Unsloth acceleration, SGLang serving, synthetic data pipelines, model merging, and standardized evaluation.

9 notebooks →

60ORPO: Reference-Free Alignment 61KTO: Binary Feedback Alignment 62GRPO: Reasoning Model Training 63Preference Algorithm Comparison 64LM Evaluation Harness 65Unsloth: 2× Faster Fine-Tuning 66SGLang: Production Inference 67Synthetic Data Generation 68Model Merging with MergeKit