experimental model to expose arco to some reasoning

after some research i notice i was finetuning models with super high lr, further models should be better since will maintain most of the power of arco

This table presents the extracted scores in a clear, tabular format. The "Task" column shows the name of each benchmark, the "Score" column displays the corresponding value, and the "Metric" column indicates whether the score is acc_norm or acc.

format is this:

Instruction: <your instruction>
Reasoning: // starting from here, the model will start to generate the resoning and output
Output:

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.