build-small-hackathon/oneiros

✦ ONEIROS

Your dreams, mapped with a small model you control — not sent to ChatGPT.

Oneiros mengubah catatan mimpi menjadi peta SVG plus panel ringkasan entitas, menggunakan Qwen2.5-7B-Instruct (GGUF) via llama-cpp-python — tanpa API LLM pihak ketiga.

Proyek ini dibuat untuk Build Small Hackathon (Gradio × Hugging Face), track An Adventure in Thousand Token Wood.

Status repositori (Day 1)

Deploy HF Space

Space resmi: https://huggingface.co/spaces/build-small-hackathon/oneiros

1. 1.Sebelum push: ./scripts/prepare_space_requirements.sh (salin wheel Linux ke requirements.txt).
2. 2.Push ke org build-small-hackathon/oneiros.
3. 3.Variables (Settings): N_GPU_LAYERS=0, N_CTX=4096; ONEIROS_SKIP_WARMUP=1 sampai preload selesai.
4. 4.Preload 2 shard Q4KM (~15–45 menit). Cek log: [oneiros] diagnosis → shard_pair_ok: True.
5. 5.Setelah Running: uji 1 mimpi di UI.

Detail: docs/08-deploy-hf-space.md · Checklist Day 2: docs/16-checklist-sebelum-day2.md.

Quick start (lokal)

Pakai `requirements-local.txt` — jangan requirements.txt (itu untuk Space Linux).

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-local.txt

# Model — opsi cepat (Q2_K ~3GB) atau Q4_K_M (2 shard)
hf download Qwen/Qwen2.5-7B-Instruct-GGUF qwen2.5-7b-instruct-q2_k.gguf --local-dir ./models

# Atau Q4_K_M (kualitas lebih baik):
# hf download Qwen/Qwen2.5-7B-Instruct-GGUF \
#   qwen2.5-7b-instruct-q4_k_m-00001-of-00002.gguf \
#   qwen2.5-7b-instruct-q4_k_m-00002-of-00002.gguf \
#   --local-dir ./models

CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python --force-reinstall --no-cache-dir

python scripts/verify_day1.py
python scripts/smoke_model.py
python tests/test_extractor.py
python scripts/run_mimpi_uji.py
python app.py

Day 1 — perintah verifikasi

Dokumentasi

Konteks Cursor: `ONEIROS_CURSOR_CONTEXT.md`.

Lisensi

TBD — tentukan sebelum publish ke Hub.

Build Small Hackathon 2026 · Oneiros