Spaces:
Running on Zero
Running on Zero
| title: Tessera-1B-Nano comparison | |
| emoji: 🧩 | |
| colorFrom: gray | |
| colorTo: blue | |
| sdk: gradio | |
| app_file: app.py | |
| suggested_hardware: zero-a10g | |
| license: apache-2.0 | |
| short_description: Side-by-side generation with a live concept readout | |
| # Tessera-1B-Nano vs Tessera-1B-Nano-Base | |
| Side-by-side demo for the token-matched comparison described in the | |
| [Tessera-1B-Nano repository](https://github.com/yava-code/Tessera-1B-Nano). Three generations per run: | |
| the NTP-only baseline, the concept arm with predicted feedback, andthe concept arm with its concept feedback zeroed at inference. A readout panel reports the teacher-forced NTP | |
| loss of each model on the prompt and the zero-feedback penalty of the concept arm. | |
| Runs on ZeroGPU: the checkpoints load on CPU at startup and each request moves compute to | |
| a shared A100 through the `spaces.GPU` decorator. | |
| The checkpoints: [Tessera-1B-Nano](https://huggingface.co/yava-code/Tessera-1B-Nano) and | |
| [Tessera-1B-Nano-Base](https://huggingface.co/yava-code/Tessera-1B-Nano-Base). The full | |
| study, including the committed intervention numbers (zeroing costs +0.1050 nats on held-out | |
| data, shuffling costs +0.0008), is in the repository whitepaper. | |