arco consistently outperforms every sota model below 600m parameters as well as some 1b base models on average on 5 core benchmarks and is competitive with the best 0.7b-1b llms. arco is a merge of multiple internal models fine-tuned on a diverse set of styles and finally merged with the several models (including palmer-004-turbo and danube3-chat), followed by a merge with base model to preserve knowledge.
prompt
there is no prompt intentionally set but this one worked really good for me:
The following is a conversation between a super smart AI assistant and an user.
user: <your question>
assistant:
benchmarks
zero-shot evaluations performed on current sota ~0.5b models.
| Parameters | Model | MMLU | ARC-C | HellaSwag | PIQA | Winogrande | Average |
|---|---|---|---|---|---|---|---|
| 0.5b | qwen2 | 44.13 | 28.92 | 49.05 | 69.31 | 56.99 | 49.68 |
| 1.1b | tinyllama | 25.77 | 30.29 | 59.35 | 73.29 | 59.59 | 49.66 |
| 0.5b | danube3-base | 24.81 | 36.18 | 60.46 | 73.78 | 61.01 | 51.25 |
| 0.5b | danube3-chat | 25.54 | 36.26 | 60.72 | 74.32 | 61.40 | 51.64 |
| 0.5b | palmer-004-turbo | 27.36 | 35.58 | 61.79 | 73.67 | 61.17 | 51.91 |
| 0.5b | arco | 26.17 | 37.29 | 62.88 | 74.37 | 62.27 | 52.60 |
supporters
trivia
arco comes from spanish word "bow" which is always associated with arrows and hence, speed.