Auditing Preference Biases and Fine-Tuning Language Models with Direct Preference Optimization on Anthropic HH-RLHF Using TRL and LoRA

In this tutorial, we design an end-to-end preference-learning workflow using the Anthropic HH-RLHF dataset and Direct Preference Optimization (DPO). We begin by preparing a robust Colab environment, loading and parsing chosen–rejected response pairs, and auditing the dataset for structural and length-based preference biases. We then run lexical shortcut diagnostics to determine whether surface-level linguistic patterns can separate preferred from rejected responses, prepare conversational data with tokenizer-aware length filtering, and construct a version-robust DPO training pipeline with TRL and optional LoRA adaptation. Finally, we fine-tune a Qwen2.5-0.5B-Instruct model, evaluate reward accuracy and training behavior, analyze performance across individual HH-RLHF subsets, inspect potential length bias, generate sample responses, and save the resulting policy for further experimentation.

We set up the required libraries, handle dependency compatibility issues, and configure the main parameters used throughout the tutorial. We also initialize reproducibility settings and inspect the available hardware, precision modes, and installed TRL interfaces. This gives us a stable environment before we process the HH-RLHF dataset and train the preference model.

We load samples from the different Anthropic HH-RLHF subsets and create balanced training and testing datasets. We parse each conversation into structured user and assistant messages while ensuring that chosen and rejected responses share the same conversational prefix. We then filter invalid pairs so that we work only with properly aligned preference examples.

We analyze the preference pairs to measure differences in response length, conversation depth, and source-specific behavior. We also train a TF-IDF and logistic regression diagnostic to test whether simple lexical patterns can distinguish chosen responses from rejected ones. This helps us detect shortcuts that the language model could potentially exploit instead of learning the intended preference signal.

We prepare the tokenizer, apply the conversational chat template, and calculate token lengths for every preference pair. We filter examples that exceed our prompt or total sequence limits and dynamically construct DPO configuration arguments based on the installed TRL version. We then load the base model, configure LoRA when available, and build the DPO trainer that we use for fine-tuning.

We train the model using Direct Preference Optimization with the configured batch size, gradient accumulation, learning rate, and optimization steps. We evaluate the resulting policy on held-out preference pairs and inspect metrics such as loss, reward margins, and reward accuracy. We also visualize the training history to observe how preference-learning performance changes throughout optimization.