base_model:
- appvoid/palmer-002-32k
- raidhon/coven_tiny_1.1b_32k_orpo_alpha
- appvoid/palmer-003
library_name: transformers
tags:
- mergekit
- merge
license: apache-2.0
Introducing palmer-004
We are thrilled to unveil palmer-004, the fourth iteration in our esteemed palmer series, designed to push the boundaries of performance and context window capacity. Merging the best of palmer-003 with the expansive capabilities of palmer-002-32k, palmer-004 is a testament to the power of merging similar trained models, think of it as palmer-003 with 32k context with minimal performance degradation.
| Model | MMLU | ARC-C | HellaSwag | PIQA | Winogrande |
|---|---|---|---|---|---|
| tinyllama-3t | 0.2577 | 0.3029 | 0.5935 | 0.7329 | 0.5959 |
| palmer-003 | 0.2523 | 0.3439 | 0.6208 | 0.7524 | 0.6590 |
| palmer-004 | 0.2601 | 0.3456 | 0.6138 | 0.7443 | 0.6511 |
Crafted using the TIES merge method, palmer-004 elevates its predecessor's prowess to a remarkable 32k context window without needing to train on additional data. This model doesn't just aim to match; it surpasses, outperforming the current state-of-the-art coven_tiny_1.1b_32k_orpo_alpha on several benchmarks. As of June 1, 2024, palmer-004 stands as the 2nd best 1b overall model as well as the best 32k 1B model to fine-tune from on Hugging Face, setting a new standard for excellence in the open-source AI community.
If you are using Windows, you can now support me by purchasing a copy of quark, a cool text editor that uses language models. Still in beta but super cool still.