appvoid / palmer-002

palmer

a better base model

palmer is a series of ~1b parameters language models fine-tuned to be used as base models instead of using custom prompts for tasks. This means that it can be further fine-tuned on more data with custom prompts as usual or be used for downstream tasks as any base model you can get. The model has the best of both worlds: some "bias" to act as an assistant, but also the abillity to predict the next-word from its internet knowledge base. It's a 1.1b llama 2 model so you can use it with your favorite tools/frameworks.

evaluation 🧪

note that this is a zero-shot setting as opposite to open llm leaderboard's few-shot evals

   Model           ARC_C   HellaSwag  PIQA  Winogrande Average
tinyllama-2      | 0.2807 | 0.5463 | 0.7067 | 0.5683 | 0.5255 |
palmer-001	     | 0.2807 | 0.5524 | 0.7106 | 0.5896 | 0.5333 |
babbage-001      | 0.2944 | 0.5448 | 0.7410 | 0.5935 | 0.5434 |
deacon-1b        | 0.2944 | 0.5727 | 0.7040 | 0.5801 | 0.5434 |
tinyllama-2.5    | 0.3191 | 0.5896 | 0.7307 | 0.5872 | 0.5566 |
palmer-002       | 0.3242 | 0.5956 | 0.7345 | 0.5888 | 0.5607 |
babbage-002      | 0.3285 | 0.6380 | 0.7606 | 0.6085 | 0.5839 |

This model shows exceptional performance and as of now is the best tinyllama-size base model. Furthermore, this proves LIMA paper point and serves as a good open-source alternative to openai's babbage-002 .

training 🦾

Training took ~3.5 P100 gpu hours. It was trained on 15,000 gpt-4 shuffled samples. palmer was fine-tuned using lower learning rates ensuring it keeps as much general knowledge as possible.

prompt 📝

no prompt 🚀

Choose this if you prefer a base model without too much fine-tuning.