Objectverse Diary Architecture

Multi-layered small model pipeline for the Build Small Hackathon (An Adventure in Thousand Token Wood)

Hugging Face Space Runtime Environment ZeroGPU Allocation Sandbox HTTPS WS/Post @zero_gpu Local GGUF Local Save gated pull GGUF pull Users Upload Image Chat Session Web/Mobile Gradio Web UI app.py / src/ui/ • Image Drag-Drop • Persona Selector • Typewriter Diary • Character Chat • SVG Card Render Example Gallery Cache (Deterministic Baseline) Pipeline Core src/pipeline.py • State Routing • Fallback Logic • Parse Schemas Pydantic validation Vision Backend src/models/vision_runner MiniCPM-V 2.6 (8B) ZeroGPU compatible Fallback to mock on failure Text Backend llama_cpp_runner.py Qwen 1.5B GGUF Merged LoRA v2 Adapter Deterministic fallback runtime Output Services renderer/ & traces/ • Card HTML Gen • Anonymizer Traces • data/traces/*.jsonl HF Hub Remote Assets • SFT Dataset • LoRA Weights • GGUF Files • Gate models Legend UI Layer (Gradio) Controller Layer (Python) Vision Engine (MiniCPM-V) Text Engine (Llama.cpp GGUF) External & File Outputs Security/Dynamic Hardware Group

UI & Frontend Layer

  • • English-First Copy: Retro archive design, warm paper, mystery vibe
  • • Deterministic Fallback: Example gallery reads committed mock records
  • • Interactive Sandbox: Full chat session maintaining object persona

Pipeline Coordinator

  • • Modular Routing: Vision descriptions trigger first-person diaries
  • • Pydantic Validation: Strict checks on JSON output from LLM
  • • Trace Compliance: Generates anonymized session logs to JSONL files

Dual-Engine Model Execution

  • • MiniCPM-V 2.6 (8B): Runs via HF Spaces ZeroGPU dynamically
  • • Llama.cpp (1.5B): Runs highly optimized GGUF adapter locally
  • • Flexible Mock Fallback: Ensures 100% runtime uptime for judges

Objectverse Diary • Created for Build Small Hackathon • June 2026