| title: Token Routing Interactive Paper | |
| emoji: 🧭 | |
| colorFrom: purple | |
| colorTo: green | |
| sdk: static | |
| app_file: index.html | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Interactive report for the final TR-HASH-0.5B. | |
| # TR-HASH-0.5B — Interactive Architecture Report | |
| Interactive companion for the final **TR-HASH-0.5B** research checkpoint. | |
| The Space provides: | |
| - an architecture explorer for the shared SwiGLU and fixed top-2 residual path; | |
| - checkpoint-derived tokenizer IDs and expert assignments for selected tokens; | |
| - the recorded 20B-token pretraining and one-epoch SFT diagnostics; | |
| - live chat generation from `Pacific-i64/TR-HASH-MOE-500M-HF`. | |
| Its metrics describe only the final 492.1M model. They are training | |
| diagnostics, not a dense-baseline comparison or a general-capability benchmark. | |
| ## Sources | |
| - [TR-HASH-MOE-500M-HF](https://huggingface.co/Pacific-i64/TR-HASH-MOE-500M-HF) | |
| - [Training framework](https://github.com/Complexity-ML/complexity-framework) | |
| - [vllm-i64](https://github.com/Complexity-ML/vllm-i64) | |