Finetuning Resistance Research

Overview

This project investigates whether "Inverse Scaling" datasets are resistant to model finetuning. We focused on the redefine-math task, where models must ignore mathematical priors to perform textual operations.

Key Findings

• Not Finetuning Proof: Fine-tuning TinyLlama-1.1B improved accuracy from 46% (Zero-shot) to 80% (LoRA).
• Stronger than ICL: Fine-tuning significantly outperformed 5-shot in-context learning (57%).
• Conclusion: Sufficient gradient updates can override strong pre-trained priors.

Reproducing Results

1. Setup:

   uv pip install torch transformers peft datasets bitsandbytes accelerate scikit-learn matplotlib pandas

2. Run Experiment:

   python run_finetuning_experiment.py

   (Note: This script now includes the robust evaluation logic)

3. Verify: Check results/redefine_math_robust.json.

Structure

• run_finetuning_experiment.py: Main training script.
• evaluate_robust.py: Evaluation script using log-likelihoods.
• datasets/: Inverse scaling data.
• results/: Checkpoints and JSON logs.