Linear models offer a promising approach to significantly reduce computational costs at scale, particularly for large context lengths. Enabling a >1000x improvement in inference costs, enabling o1 inference time thinking and wider AI accessibility.
As demonstrated with our Qwerky-72B-Preview and prior models such as QRWKV6-32B Instruct Preview, we have successfully converted Qwen 2.5 72B into a RWKV variant without requiring a pretrain on the base model or retraining the model from scratch. Enabling us to test and validate the more efficient RWKV Linear attention with a much smaller budget. Since our preview, we have continued to refine our technique and managed to improve the model over the preview model iteration.
As with our previous models, the model's inherent knowledge and dataset training are inherited from its "parent" model. Consequently, unlike previous RWKV models trained on over 100+ languages, the QRWKV model is limited to approximately 30 languages supported by the Qwen line of models.
You may find our details of the process from our previous release, find it here.
Benchmarks is as follows for both Qwerky-QwQ-32B and Qwerky-72B models:
| Tasks | Metric | Qwerky-QwQ-32B | Qwen/QwQ-32B | Qwerky-72B | Qwen2.5-72B-Instruct |
|---|---|---|---|---|---|
| arc_challenge | acc_norm | 0.5640 | 0.5563 | 0.6382 | 0.6323 |
| arc_easy | acc_norm | 0.7837 | 0.7866 | 0.8443 | 0.8329 |
| hellaswag | acc_norm | 0.8303 | 0.8407 | 0.8573 | 0.8736 |
| lambada_openai | acc | 0.6621 | 0.6683 | 0.7539 | 0.7506 |
| piqa | acc | 0.8036 | 0.7976 | 0.8248 | 0.8357 |
| sciq | acc | 0.9630 | 0.9630 | 0.9670 | 0.9740 |
| winogrande | acc | 0.7324 | 0.7048 | 0.7956 | 0.7632 |
| mmlu | acc | 0.7431 | 0.7985 | 0.7746 | 0.8338 |
Note: All benchmarks except MMLU are 0-shot and Version 1. For MMLU, it's Version 2.
- Downloads last month
- 21