AI Engineer Needed: Fix LLM Fine-Tuning Bug / Troubleshooting

We are looking for an experienced AI Engineer or Data Scientist for a quick, micro-project to debug and resolve a specific issue in our LLM fine-tuning pipeline. This is a straightforward troubleshooting task for an expert who knows training loops inside and out. We need you to look at our setup/logs, identify the bottleneck, and implement a clean fix so we can resume successful training. Common issues we might be facing (or similar): * Formatting/tokenization issues or train/serve mismatches. * Training loop anomalies (loss spikes, overfitting, catastrophic forgetting). * Hardware/memory optimization (CUDA OOM during validation/training). * PEFT/LoRA adapter loading or weight merging bugs. Requirements: * Hands-on experience fine-tuning open-source LLMs (Llama, Mistral, Qwen, etc.). * Deep familiarity with frameworks like Hugging Face (TRL/SFTTrainer), Unsloth, Axolotl, or native PyTorch. * Expert-level debugging skills in Python and CUDA environments. * Ability to deliver a working fix quickly.