metadata
license: llama2
The LLaMA-2-7b model finetuned on the Math task using CorDA in the IPA mode with MetaMath.
| Method | TriviaQA | NQ open | GSM8k | Math |
|---|---|---|---|---|
| LoRA | 44.17 | 1.91 | 42.68 | 5.92 |
| CorDA (KPA with nqopen) | 45.23 | 10.44 | 45.64 | 6.94 |
| CorDA (IPA with MetaMath) | - | - | 54.59 | 8.54 |
You can evaluate the model's performance following the step-3 in CorDA github repo.
Note: The model trained using CorDA adapter is based on customized code. If you want to restore the original LLaMA architecture, execute merge_adapter_for_corda.py in CorDA github repo.