RLHF Specialist

Responsibilities:

• Generate high-quality preference data by comparing multiple model responses and ranking them based on criteria such as helpfulness, honesty, and harmlessness (HHH).
• Design complex, multi-turn prompts to stress-test model behavior and expose weaknesses in reasoning or safety.
• Write detailed “chain-of-thought” explanations and rationales to train reward models on why specific responses are superior.
• Collaborate with Machine Learning Engineers to analyze model failure modes and identify data gaps that, when filled, will improve reinforcement learning outcomes.
• Develop and iterate on annotation strategies for preference scoring and reinforcement signals, ensuring consistency across a global team.
• Proactively probe models to identify vulnerabilities, biases, or hallucination patterns, documenting findings for model optimization.
• Analyze edge cases where the reward model behaves unexpectedly (e.g., over-indexing on verbosity or style over substance). Provide detailed feedback to ML engineers on reward model failure modes and suggest specific data interventions to correct model behavior.
• Develop and document templated instruction sets for larger annotation teams. Translate complex reinforcement learning concepts into simple, repeatable tasks for junior reviewers, ensuring high-quality data collection at scale.
• Monitor model performance over time by maintaining a personal test set of prompts.
• Regularly re-evaluate new model versions against historical benchmarks to track improvements or regressions in reasoning and alignment.

Requirements:

• Minimum of 4  years of experience in Data Annotation, Model Evaluation, Computational Linguistics, or Trust and Safety, specifically working with AI/ML training data.
• Strong proficiency in Python and deep learning frameworks (PyTorch, JAX, or TensorFlow).
• Deep understanding of Reinforcement Learning concepts (PPO, Trust Regions, Reward Hacking) and how they apply to language generation.
• Hands-on experience fine-tuning open-source models (e.g., Llama 2/3, Mistral, gemma) using techniques like LoRA/QLoRA.
• Experience working with annotation tools (LabelBox, Scale AI, Snorkel) and managing human-in-the-loop workflows.
• Ability to diagnose why an RL policy collapsed and adjust hyperparameters or reward structure accordingly.
• Experience with Constitutional AI or Self-Alignment techniques.
• Contributions to open-source alignment libraries (TRL, Transformer Reinforcement Learning, Axolotl).
• Experience with cloud Platforms (AWS SageMaker, GCP Vertex AI).

Remuneration: NGN 500,000 monthly