RM -RF: Official Repo for SANER 2026 Paper "RM -RF: Reward Model for Run-Free Unit Test Evaluation"

This repository contains the data used in the experiments:

• The file holdout_dataset.zip includes three .jsonl files — one for each target type:
  binary_type, float_type, and reverse_binary_type.

• The file train_data.zip contains six .jsonl files: training (train) and validation (val) splits for each of the three target types.

• The validation_subset directory contains .toml files — one per sample.

• The slurm_scripts_examples directory contains example .sh scripts for:

  • Full fine-tuning,
  • Fine-tuning with LoRA,
  • Inference.

• The prompts directory includes:

  • A prompt template used to generate test-breaking inputs,
  • Descriptions of errors that were synthetically generated using an LLM.