Spaces:
Running on Zero
Running on Zero
fix hf-mount init failure: drop models: pre-mount (training repo is large and growing; app pulls only what it needs at runtime)
e8b14b9 verified | title: Beatrix — AlephLLM Chat | |
| emoji: 🐠 | |
| colorFrom: yellow | |
| colorTo: yellow | |
| sdk: gradio | |
| sdk_version: 6.23.1 | |
| python_version: '3.12' | |
| app_file: app.py | |
| pinned: false | |
| license: mit | |
| short_description: Talk to Beatrix — a live byte-level AlephLLM training run | |
| # Beatrix — AlephLLM chat | |
| A live window onto the AlephLLM training runs: this space always serves | |
| the **newest checkpoint** of `mini-beatrix-1` (112.5M params, byte-level, | |
| currently pretraining) straight from the | |
| [training repo](https://huggingface.co/AbstractPhil/alephllm-mini-beatrix-training). | |
| - **Completion tab** — raw next-byte continuation, streamed. | |
| - **Chat tab** — a transcript-format preview (the model has had no chat | |
| training yet; a chat-tuned stage lands after base pretraining). | |
| - **Reload** — pulls the latest checkpoint mid-run; Beatrix improves as | |
| the run progresses. | |
| The model reads raw UTF-8 bytes through a trigram-composed embedding and | |
| routes through signed geometric addresses (the aleph mechanism — | |
| `sinh/Σcosh` dispatch, no softmax-over-choices, inhibition first-class), | |
| with expert banks and an address-based head read that are **born at | |
| exactly zero** and must earn their way in by gradient. | |
| Code and training stack: [github.com/AbstractEyes/alephllm](https://github.com/AbstractEyes/alephllm) | |