🧠 Alignment of Large Language Models with Constrained Learning

This repo contains the implementation for the submission: "Alignment of Large Language Models with Constrained Learning".

---

📁 Directory Structure

algorithms/ – Dual Optimization Algorithms

• dual_optimizer.py: Implements the dual variable computation algorithm.

data_prep/ – Data Preparation Scripts

• dataset_configs/: Custom dataset class definitions.
• collect.py: Collects prompt datasets.
• generate.py: Generates responses using a specified model and prompt dataset.
• score.py: Evaluates responses using reward and cost models.
• templates/: Prompt templates used for model-based generation.

trainers/ – Training Scripts

• multi_shot_trainer.py: Multi-shot training loop.
• one_shot_trainer.py: One-shot training loop.

gpt_based/ – GPT-Based Evaluation Scripts

• adversarial_collect.py: Collects adversarial prompts.
• gpt_generate.py: Generates responses for GPT-based evaluation.
• gpt_eval.py: Evaluates responses using a GPT model as the evaluator.
• templates/: Prompt templates for GPT-based evaluation.

visualize/ – Plotting Scripts

• Scripts to generate all figures presented in the paper.

exec/ – Execution Scripts

• Bash scripts for running the scripts.

---