rui-ye / openfedllm Goto Github PK

Hello, directly run your code, the configuration has not changed, the test steps are also according to your evaluation/open_ended process, but gen_judge_mtbench.py results are all ERROR. What's the problem? Looking forward to your reply.

Run the following command:

1. sh training_scripts/run_sft.sh
2. python utils/merge_lora.py --base_model_path "meta-llama/Llama-2-7b-hf" --lora_path "output/alpaca-gpt4_20000_fedavg_c20s2_i10_b16a1_l512_r32a64_20240313104822/checkpoint-200"
3. CUDA_VISIBLE_DEVICES=2 python gen_model_answer_mt.py --base_model_path "../../output/alpaca-gpt4_20000_fedavg_c20s2_i10_b16a1_l512_r32a64_20240313104822/full-200" --template "alpaca"
4. python gen_judge_mtbench.py --judge_model gpt-4-1106-preview --model_list alpaca-gpt4_20000_fedavg_c20s2_i10_b16a1_l512_r32a64_20240313104822_200
5. python show_results_mt.py --model_list alpaca-gpt4_20000_fedavg_c20s2_i10_b16a1_l512_r32a64_20240313104822_200 --judge_model gpt-4-1106-preview

Here are the results of running on the A40:
model_answer.zip
model_judgment.zip

Thanks for your brilliant work. I would like to do SFT with multiple GPUs. Does your framework support this feature by design or I need to make some modifications?

Thx for ur good work. Could u pls tell me how can we run the local dataset? If the prompt in my dataset is same with the vicgalle/alpaca-gpt4, but it's not the hugging face format, it's the jason format. How can we run the code?

Hello and thank you for your great work. In addition to MT-Bench, Vicuna benchmark, and Advbench, do you have test details or codes for the other datasets?

I want to perform the local training without any federated learning operations in the same settings. So how can I modify the parameters in the off-the-shelf code framework?

I have generated several checkpoint model folders after 200 rounds of training, can I continue training with this model next time? If so, how to use the trained model parameters.

Congratulations on your excellent work！
Currently, the evaluation scripts only include open-ended scripts. Could you also release the closed-ended and other test dataset evaluation scripts? Due to the different settings (such as few-shot numbers and templates) used for different datasets, the results can vary significantly. To facilitate better comparisons and support the community's ongoing work, it would be beneficial to make this part of the code open-source.

Thanks to the authors for contributing OpenFedLLM to the open source community. Currently, I finished training the model using the default run_sft.sh script, and save to model to the output file, checkpoint-XX. When I try to use the evaluation script <gen_model_answer_mt.py> and loading the checkpoint path, I'm not able to run due to some config error
(Can't load the configuration of <file_path>, If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure <file_path> is the correct path to a directory containing a config.json file).

Any help or advice?

Hello, is this train_loss so high?

The summation of L2 losses over named_parameters() does not equal the L2 loss of the entire model. Please check fed_local_dpo.py and fed_local_sft.py