To train the models, you must first deploy an inference acceleration service and record its node IP. Then, run the corresponding training script.
For DS-R1-7B:
# Start the inference accelerator and record the node IP (e.g., 225.13.1.4)
bash ./train_script/train_7B_1_vs_1/7b_parameter_serve.sh
# Begin training using the recorded IP
SERVE_NODE_IP='225.13.1.4' bash ./train_script/train_7B_1_vs_1/max_step_2000_eval_step_32_init_1_vs_1.sh For DS-R1-14B:
# Start the inference accelerator and record the node IP (e.g., 225.13.1.4)
bash ./train_script/train_14B_1_vs_1/14b_parameter_serve.sh
SERVE_NODE_IP='225.13.1.4' bash ./train_script/train_14B_1_vs_1/max_step_2000_eval_step_32_init_1_vs_1.shEvaluation scripts for TLDR and baselines are included. To run evaluation:
bash ./eval_script/eval_tldr_weight.shbash ./eval_script/eval_tldr_weight.shWe provide the training data used in our experiments under the ./data/data_repo directory:
- ./data/data_repo/7b_train: Training data for the 7B model
- ./data/data_repo/14b_data: Training data for the 14B model
- ./data/data_repo/eval_set: Validation set data