| ---
|
| title: SICS-1-35B · Фабрика гипотез
|
| emoji: ⚗️
|
| colorFrom: blue
|
| colorTo: green
|
| sdk: gradio
|
| sdk_version: 6.14.0
|
| python_version: '3.12'
|
| app_file: app.py
|
| pinned: false
|
| ---
|
|
|
| # SICS-1-35B · «Фабрика гипотез»
|
|
|
| The public demo endpoint for **SICS-1-35B** — our fine-tuned 35B
|
| mixture-of-experts reasoning model, the engine behind **«Фабрика гипотез»**
|
| (Hypothesis Factory): an AI co-scientist that turns froth-flotation tailings
|
| analysis into ranked, cited, testable process/equipment hypotheses to recover
|
| more Ni and Cu.
|
|
|
| Served as a **GGUF `Q4_K_M`** via `llama-cpp-python` on **ZeroGPU**. Reachable
|
| programmatically via `gradio_client`.
|
|
|
| The model is lazily loaded with `Llama(n_gpu_layers=-1)` created **inside** the
|
| `@spaces.GPU` call, so CUDA is initialised in the GPU worker, not the main
|
| process — which is what ZeroGPU requires.
|
|
|
| Panels: **🖥 Превью** (live HTML/JS artifact), **🧩 Инструменты**,
|
| **🧩 Код**, **🧠 Рассуждение**, **📝 Raw**.
|
|
|
| ## Required settings
|
| - **Hardware:** ZeroGPU (`zero-a10g`) — fits the ~21 GB Q4 quant with room for KV.
|
| - **Secrets:** `HF_TOKEN` (read access to the weights repo), `GGUF_REPO`, `GGUF_FILE`
|
| — the private GGUF repo + file path, kept out of the public code.
|
|
|
| ## Tunable variables
|
| | var | meaning |
|
| |-----|---------|
|
| | `GGUF_REPO` | GGUF repo (set as a secret) |
|
| | `GGUF_FILE` | quant file to load (set as a secret) |
|
| | `N_CTX` | context window (default 16384) |
|
|
|
