reproduce_lora_exp.md

🚀 Engram vs LoRA Catastrophic Forgetting

This guide helps you reproduce the Catastrophic Forgetting experiments comparing TinyEngram and LoRA, as detailed in Key Finding 2.

🧹 Environment Setup

Please follow the same environment setup as the Quick Start Guide.

📥 Dataset Preparation

We use the glaive-function-calling-v2 dataset to create a "poison" dataset that induces a strong distribution shift (forcing function call formats).

1. Generate the Dataset: Run the provided processing script to download and filter the dataset.

# Ensure the output directory exists
mkdir -p dataset/glaive

# Run processing script
python data/process_glaive_poison.py --output dataset/glaive/glaive.parquet

This will create dataset/glaive/glaive.parquet containing only pure function-call examples.

🏋️ Training

We provide two training scripts: one for TinyEngram and one for LoRA.

1. Train TinyEngram

bash train/train_poison.sh

• Config: By default, it uses a small Engram configuration (vocab_size=1000 200, layers=1 2 3 4) designed for this experiment.
• Output: Checkpoints will be saved in output_poison_engram/.

2. Train LoRA (Baseline)

bash train/train_poison_lora.sh

• Config: Uses Rank=16 which matches the trainable parameter count of the TinyEngram configuration (~30M params).
• Output: Checkpoints will be saved in output_poison_lora_r16/.

Note: You may need to adjust MODEL path and VISIBLE_GPUS in the script files to match your local environment.

📊 Evaluation

We evaluate the trained models on TruthfulQA to measure how much general capability (truthfulness/common sense) is retained after fine-tuning on the "poison" dataset.

1. Evaluate TruthfulQA (MC1/MC2)

Go to the eval_scripts directory and run the evaluation scripts. Remember to update the MODEL_PATH inside these scripts to point to your trained checkpoints.

cd eval_scripts

# Evaluate TinyEngram
bash run_eval_truthfulqa_engram.sh

# Evaluate LoRA
bash run_eval_truthfulqa_lora.sh

# Evaluate Base Model (Optional reference)
bash run_eval_truthfulqa_base.sh

Results: The scripts will output the MC1 and MC2 scores. Compare these with the base model scores to observe the degree of catastrophic forgetting.