| # Chichu 1.5 Pro | |
| **Chichu 1.5 Pro** is a ~16M parameter language model built from scratch. | |
| Named after a cat named Chichu, created by Sebastian. | |
| ## Model Details | |
| - **Architecture**: Custom GPT-2 style Transformer | |
| - **Parameters**: ~16M | |
| - **Embedding dim**: 256 | |
| - **Layers**: 4 | |
| - **Heads**: 4 | |
| - **FFN dim**: 1024 | |
| - **Max sequence length**: 384 | |
| - **Vocab size**: 50,259 (GPT-2 + 2 special tokens: <|im_start|>, <|im_end|>) | |
| ## Training | |
| Trained on the r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation dataset | |
| (890K high-quality SFT examples distilled from Qwen3.8-max, GLM-5.2, and Kimi-K3). | |
| ## Chat Format | |
| ChatML with: `<|im_start|>user | |
| ...<|im_end|>` / `<|im_start|>assistant | |
| ...<|im_end|>` | |
| ## License | |
| MIT | |