CLASS: A Design Framework for Building Intelligent Tutoring Systems Based on Learning Science Principles (EMNLP 2023)
Shashank Sonkar, Naiming Liu, Debshila Basu Mallick, Richard G. Baraniuk
Paper: https://arxiv.org/abs/2305.13272
Branch: CLASS
Pedagogical Alignment of Large Language Models (EMNLP 2024)
Shashank Sonkar*, Kangqi Ni*, Sapana Chaudhary, Richard G. Baraniuk
Paper: https://arxiv.org/abs/2402.05000
Branch: main
This repo aims to develop effective intelligent tutoring agents that help students develop critical thinking and problems solving skills.
Please refer to scripts/run.sh as an example, which runs the training and evaluation of a selected model using 4*A100 GPUs. To run this example without training, download the models from the section below and refer to scripts/run_no-train.sh.
The following subsections break down scripts/run.sh with more detailed explanations.
The training and evaluation use bio-dataset-1.json, bio-dataset-2.json, bio-dataset-3.json, and bio-dataset-ppl.json from the datasets folder. Each contains mock conversations between a student and a tutor based on biology concepts generated from OpenAI's GPT-4. These data are then preprocessed into the required formats for training and evaluation datasets. Please refer to the branch CLASS for instructions on generating these data.
Set the user parameters:
FULL_MODEL_PATH="meta-llama/Meta-Llama-3.1-8B-Instruct"
MODEL_DIR="models"
DATA_DIR="datasets"
SFT_OPTION="transformers" # choices: ["transformers", "fastchat"]
ALGO="dpo" # choices: ["dpo", "ipo", "kto"]
BETA=0.1 # choices: [0.0 - 1.0]
Preprocess data:
python src/preprocess_sft_data.py --data_dir $DATA_DIR
We provide 2 options for SFT: (1) Transformers (2) FastChat.
(1) Run SFT with Transformers:
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=20001 src/train/train_sft.py \
--model_path $FULL_MODEL_PATH \
--train_dataset_path $SFT_DATASET_PATH \
--eval_dataset_path ${DATA_DIR}/bio-test.json \
--output_dir $SFT_MODEL_PATH \
--cache_dir cache \
--bf16 \
--num_train_epochs 3 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 2 \
--evaluation_strategy "epoch" \
--eval_accumulation_steps 50 \
--save_strategy "epoch" \
--seed 42 \
--learning_rate 2e-5 \
--weight_decay 0.05 \
--warmup_ratio 0.1 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--max_seq_length 4096 \
--gradient_checkpointing
(2) Run SFT with FastChat:
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=20001 FastChat/fastchat/train/train.py \
--model_name_or_path $FULL_MODEL_PATH \
--data_path $SFT_DATASET_PATH \
--eval_data_path ${DATA_DIR}/bio-test.json \
--output_dir $SFT_MODEL_PATH \
--cache_dir cache \
--bf16 True \
--num_train_epochs 3 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 2 \
--evaluation_strategy "epoch" \
--eval_accumulation_steps 50 \
--save_strategy "epoch" \
--seed 42 \
--learning_rate 2e-5 \
--weight_decay 0.05 \
--warmup_ratio 0.1 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--tf32 True \
--model_max_length 4096 \
--gradient_checkpointing True
Generate preference data:
CUDA_VISIBLE_DEVICES=0,1,2,3 python src/evaluate/generate_responses.py --model_path $SFT_MODEL_PATH --output_dir ${SFT_MODEL_PATH}/final_checkpoint-dpo --test_dataset_path $DPO_DATASET_PATH --batch_size 256
python src/preprocess/preprocess_dpo_data.py --response_file ${SFT_MODEL_PATH}/final_checkpoint-dpo/responses.csv --data_file $DPO_PREF_DATASET_PATH
Run preference alignment:
DPO_MODEL_PATH="${MODEL_DIR}_dpo/${MODEL_NAME}_bio-tutor_${ALGO}"
CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch --config_file=ds_config/deepspeed_zero3.yaml --num_processes=4 train/train_dpo.py \
--train_data $DPO_PREF_DATASET_PATH \
--model_path $SFT_MODEL_PATH \
--output_dir $DPO_MODEL_PATH \
--beta $BETA \
--loss $ALGO \
--gradient_checkpointing \
--bf16 \
--gradient_accumulation_steps 4 \
--per_device_train_batch_size 2 \
--num_train_epochs 3
Evaluate the accuracy and F1 scores of SFT and Aligned models:
# Generate responses from the SFT model
CUDA_VISIBLE_DEVICES=0,1,2,3 python src/evaluate/generate_responses.py --model_path $SFT_MODEL_PATH --output_dir ${SFT_MODEL_PATH}/final_checkpoint-eval --test_dataset_path $TEST_DATASET_PATH --batch_size 256
# Generate responses from the Aligned model
CUDA_VISIBLE_DEVICES=0,1,2,3 python src/evaluate/generate_responses.py --model_path $DPO_MODEL_PATH --output_dir ${DPO_MODEL_PATH}/final_checkpoint-eval --test_dataset_path $TEST_DATASET_PATH --batch_size 256
# Evaluate the SFT model
echo "Metrics of the SFT Model:"
python src/evaluate/evaluate_responses.py --response_file ${SFT_MODEL_PATH}/final_checkpoint-eval/responses.csv
# Evaluate the Aligned model
echo "Metrics of the RL Model:"
python src/evaluate/evaluate_responses.py --response_file ${DPO_MODEL_PATH}/final_checkpoint-eval/responses.csv
Evaluate the ppl of SFT and Aligned models:
CUDA_VISIBLE_DEVICES=0,1 python src/evaluate/evaluate_ppl.py --model_path $SFT_MODEL_PATH
CUDA_VISIBLE_DEVICES=0,1 python src/evaluate/evaluate_ppl.py --model_path $DPO_MODEL_PATH
For easier access to the models, download them from Hugging Face.
SFT Models:
- Llama-3.1-8B-Instruct_bio-tutor_sft
- Mistral-7B-Instruct-v0.2_bio-tutor_sft
- zephyr-7b-beta_bio-tutor_sft
Aligned Models:
- Llama-3.1-8B-Instruct_bio-tutor_dpo
- Mistral-7B-Instruct-v0.2_bio-tutor_dpo
- zephyr-7b-beta_bio-tutor_dpo
- Llama-3.1-8B-Instruct_bio-tutor_kto
- Mistral-7B-Instruct-v0.2_bio-tutor_kto
- zephyr-7b-beta_bio-tutor_kto
If you find our work useful, please cite:
@misc{sonkar2023classdesignframeworkbuilding,
title={CLASS: A Design Framework for building Intelligent Tutoring Systems based on Learning Science principles},
author={Shashank Sonkar and Naiming Liu and Debshila Basu Mallick and Richard G. Baraniuk},
year={2023},
eprint={2305.13272},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2305.13272},
}
@misc{sonkar2024pedagogical,
title={Pedagogical Alignment of Large Language Models},
author={Shashank Sonkar and Kangqi Ni and Sapana Chaudhary and Richard G. Baraniuk},
year={2024},
eprint={2402.05000},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2402.05000},
}