| Lychee-27B |
|
|
| Derived from Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, which is licensed under the |
| Apache License, Version 2.0 (LICENSE.apache-2.0). |
|
|
| Modifications: the weights were updated by fine-tuning. A linear readout over 255 single-token answer codes |
| (readout.safetensors) replaces the generation output layer. At inference the full-attention layers read the whole |
| input (only padding is masked), and the answer is read from the last position; nothing is generated. |
|
|
| The modifications and this repository's code are licensed under CC BY-NC 4.0 (LICENSE). No affiliation with or |
| endorsement by Qwen or Alibaba Cloud is implied. |
| |
