CODE for PEFT using TNO with SB

1. Environments:

Pytorch >= 2.5.1
transformers >= 4.44.2
huggingface-hub >= 0.27.0
*numpy may need to be downgraded to 1.26.4

2. Additional packages:
Please install the additional packages using the following script:

pip install wandb datasets accelerate sentencepiece opt_einsum

3. Quick Start
a. Run the following script as a background agent process (recommended using screen terminal multiplexer):

bash run_agent.sh

b. Run the following script in a new terminal (recommended using screen terminal multiplexer):

bash run_tno_sb.sh

4. Reproduce of the result of SIQA with Params 0.024% (Acc: 75.1).

Run the following script:

bash run_siqa_demo.sh

5. Notes: 
a. The real-time log file for SB process can be found in center_log founder
b. If the agent process crashes unexpectedly or need to change the GPU device number, please restart will the following process:
   1. empty the folder agent_pool and restart the agent (bash run_agent.sh).
   2. in the *.POOL file in agent_pool, input the dataset name (should be the same as that in run_qtn_sb.sh) and save.
c. Each time before running a new experiment, please empty the folders: agent_pool, job_pool and result_pool.



