Invesitgating Trace-based Knowledge Distillation

Installation

# Clone the repository
conda create -n trace_kd python=3.10
<<<<<<< HEAD
cd trace_kd
=======
conda activate trace_kd
cd Trace_Check_QA
>>>>>>> 525620a5061f1b7a2c8f7a55e0c8d6a4ea51fff6
pip install -r requirements.txt
cd ..

Additional step if CUDA Toolkit is missing

conda install -c conda-forge cudatoolkit-dev -y

Usage

1. Run inference on CotempQA dataset:

# example
python inference.py \
    --model_name meta-llama/Llama-3.2-1B \
    --data_path data/cotempqa/mix.json \
    --mode default \
    --output_dir results/Cotempqa/evaluation_outputs/ \
    --evaluate_result_dir results/Cotempqa/evaluation_results/

2. Create SFT dataset for CotempQA with Input/Output format (ensure CSV is formatted correctly):

# example
python create_sft_dataset.py \
    --csv_path data/cotempqa/dataset_with_labels.csv \
    --input_col question \
    --output_col answer \
    --output_dir data/cotempqa/sft_dataset

3. Create SFT dataset for CotempQA with Input/Reasoning/Output format (ensure CSV is formatted correctly):

# example
python create_sft_dataset.py \
    --csv_path data/cotempqa/dataset_with_labels.csv \
    --input_col question \
    --output_col answer \
    --output_dir data/cotempqa/sft_dataset_with_reasoning \
    --include_reasoning \
    --reasoning_col label

Push data to your Huggingface and load that dataset in the SFT scripts accordingly. (Data will be made public later on.)

4. SFT on Cotempqa example using default settings with QLoRA for the CotempQA SFT dataset (input/output)

sbatch ./scripts/bash_scripts/cotempqa_sft_vanilla.sh "meta-llama/Llama-3.2-1B-Instruct"

5. SFT on Cotempqa example using default settings with QLoRA for the CotempQA SFT dataset with Temporal Relation and Facts in reasoning trace (input/reasoning + output) (gold labels + facts)

sbatch ./scripts/bash_scripts/cotempqa_sft_reasoning_facts.sh "meta-llama/Llama-3.2-1B-Instruct"

6. Inference

# for SFT models without any reasoning trace
sbatch ./scripts/bash_scripts/cotempqa_sft_inference.sh "meta-llama/Llama-3.2-1B-Instruct" False False

# for SFT models with reasoning trace with only temporal relation
sbatch ./scripts/bash_scripts/cotempqa_sft_inference.sh "meta-llama/Llama-3.2-1B-Instruct" True True
<<<<<<< HEAD

=======

>>>>>>> 525620a5061f1b7a2c8f7a55e0c8d6a4ea51fff6