Learning to Control Summaries with Score Ranking

Code for the paper: Learning to Control Summaries with Score Ranking.

Backbones

• Llama3.1-8B
• Qwen-7B
• Mistral-7B
• GPT-4o

Main scripts

1. Produce_summary.py Generate summaries with the chosen backbone for training and evaluation.

2. Call_finesure.py Score summaries with GPT Model (e.g., GPT-4o) for training and evaluation.

3. Control_finetune.py Fine-tune the model using the scored data.

Training is an iterative loop: 1 -> 2 -> 3 -> repeat.

4.1 Evaluate_control.ipynb Control evaluation and contour plots.

4.2 Evaluate_ranking.py (uses Llama_score.py) Ranking evaluation.

Saved models

• Trained LLaMA model (LLaMA* in the paper): https://huggingface.co/hongyeeliu/Control_Summaries_LLaMA
• Best ranking model, trained Qwen (Qwen* in the paper): https://huggingface.co/hongyeeliu/Control_Summaries_Qwen

See Produce_summary.py for running control inference.