Sekoya/alpha-signal-extractor

fix(hf): use pre-built llama-cpp-python wheel to avoid 3h build

fix(hf): upgrade llama-cpp-python 0.3.10, warm-up Ollama, Streamlit deprecation

fix: use Sekoya fine-tuned model + mmproj download for HF Spaces

feat: bypass HF 1GB git limit by downloading a fallback Qwen2-VL model directly inside the container

fix: downgrade llama-cpp-python req to >=0.2.62 to prevent source compilation on HF Spaces

fix: use official abetlen pre-built wheel for llama-cpp-python to fix 404/build timeout

fix: hardcode exact whl URL for llama-cpp-python to bypass building

fix: force pip to download specific pre-compiled wheel for llama-cpp-python

fix: use pre-compiled wheels for llama-cpp-python to prevent build timeout

fix: copy all files before pip install editable

fix: install zstd required by Ollama installer

chore: add huggingface spaces configuration to README

feat: complete DDD architecture and HF deployment

docs: update architecture & reproductibility

feat: add multimodal alpha signal extractor

docs: update architecture diagrams to include mmproj.gguf

docs: add mmproj download instructions for Qwen2.5-VL vision support

feat(vlm): enable Qwen2.5-VL vision capabilities via mmproj and Qwen25VLChatHandler in llama-cpp

docs: update architecture diagram to reflect new llama-cpp VLM path

docs: update documentation to reflect llama-cpp-python integration for Qwen2.5-VL

fix: revert to llama3.2-vision (Ollama doesn't support Qwen2.5-VL arch)

feat: switch to Qwen2.5-VL 3B for fine-tuning (smaller, fits T4 GPU)

Initial commit