About the role
As a Fine-Tuning Engineer at TheWebVale, you will specialize open-weight foundation models to perform specific client business tasks. You will adapt models like Llama, Mistral, and Qwen for legal document parsing, clinic terminology triage, and customized educational tutoring.
You will partner with AI research engineers and backend developers throughout our two-to-three week release cadences. You will balance model training loss with inference speed, deploying specialized weights that clients own and run privately.
What you will do
- Prepare, format, and tokenize instruction datasets for supervised fine-tuning (SFT).
- Fine-tune open-weight models using PEFT techniques including LoRA, QLoRA, and full weights.
- Monitor training metrics, learning rate schedules, and loss curves to prevent overfitting.
- Evaluate fine-tuned models against baseline foundation weights across standard benchmarks.
- Quantize and export fine-tuned adapter weights for low-latency serving on local servers.
- Document training hyperparameters, compute budgets, and reproducibility steps.
What you will need
- Possess 3–5 years of software or machine learning experience with dedicated fine-tuning work.
- Demonstrate deep proficiency with Python, PyTorch, Hugging Face Transformers, and PEFT.
- Show practical experience fine-tuning models on GPU cloud instances with CUDA.
- Understand tokenization nuances, padding strategies, loss functions, and weight merging.
- Possess disciplined habits tracking experiments using tools like Weights & Biases or MLflow.
- Communicate model performance gains and hardware trade-offs clearly to team leaders.
Nice to have
- Experience with preference alignment methods like DPO (Direct Preference Optimization).
- Familiarity with distributed training tools like DeepSpeed, FSDP, or Unsloth.
- Knowledge of synthetic dataset generation for fine-tuning specialized tasks.