docs/DEVELOPMENT_STATUS.md

Development Status

Last updated: 2026-06-09

Completed

• Project skeleton, README, AGENTS instructions, and Gradio app entrypoint.
• Mock MVP flow: upload/description, personality mode, object JSON, persona JSON, diary, object chat, share card, and trace saving.
• Archive-style Gradio UI with English-first / Chinese-second copy and six stable examples.
• Stable demo baseline:
  • example buttons replay committed sample traces before falling back to live generation
  • cached and live generation share the same UI output formatter
  • manual upload/description path still saves new runtime traces
• Trace and dataset tooling:
  • six public mock sample traces
  • public trace JSONL export
  • deterministic SFT preview JSONL
  • initial-stage acceptance script
• Hugging Face Space created: https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary
• MiniCPM-V 2.6 optional vision backend wiring with mock fallback.
• Optional llama.cpp / llama-cpp-python text runtime wiring through TEXT_MODEL_PATH, with mock fallback.
• Space VLM validation tooling:
  • scripts/check_space_vlm.py
  • hidden /vision_runtime_probe API for non-secret MiniCPM-V diagnostics
  • probe output support in Space VLM markdown and JSON reports
  • failure-note updater for the latest Space VLM failure summary
  • failed L4 validation report at docs/SPACE_VLM_REPORT.md
  • optional --trace-output-dir evidence export for validation traces
• Hosted MiniCPM-V 2.6 ZeroGPU validation passed on 2026-06-08 for public mug, keyboard, and shoe images after adding an HF_TOKEN Space secret for gated model access.
• Public Space live vision target:
  • hardware: zero-a10g
  • OBJECTVERSE_VISION_BACKEND=minicpm-v
  • VISION_MODEL_ID=openbmb/MiniCPM-V-2_6
  • OBJECTVERSE_TEXT_BACKEND=mock
  • text generation intentionally remains mock for this release
• Real VLM trace evidence from the hosted validation is available under data/traces/space-vlm/; text generation remained mock by design.
• ZeroGPU compatibility:
  • optional src/utils/zero_gpu.py
  • Gradio generation callback wrapped with @zero_gpu(duration=180)
  • hidden /zero_gpu_probe API confirms ZeroGPU CUDA availability when run through direct hf.space URL
• Stable submission materials:
  • Field Notes draft
  • demo video script
  • social post draft
  • stable submission guide
• Well-Tuned evidence:
  • 200-row synthetic curated v2 SFT dataset published at https://huggingface.co/datasets/qqyule/objectverse-diary-sft-curated
  • Modal Qwen 1.5B LoRA v2 run completed with 120 steps, 180 train rows, and 20 eval rows
  • LoRA v2 adapter published at https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora
• LoRA v2 GGUF runtime evidence:
  • scripts/check_llama_cpp_smoke.py
  • adapter merged into Qwen/Qwen2.5-1.5B-Instruct
  • pinned llama.cpp commit: 8f83d6c271d194bde2d410145a0ce73bc42e85cd
  • published Q4_K_M GGUF: https://huggingface.co/qqyule/objectverse-diary-qwen15b-lora/blob/main/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf
  • local smoke passed with llama-cpp text generation, schema-valid persona/diary, non-empty chat, and no text-fallback-to-mock
  • trace runtime no longer records literal TEXT_MODEL_PATH
• Local tests and initial acceptance currently pass.

Not Completed

• Hosted Space text runtime validation with the published GGUF. The public Space still uses mock text until this passes.
• Real text model traces from the hosted non-mock text runtime.
• Published Field Notes URL, recorded demo video URL, social post URL, and final public submission.

Current Runtime Defaults

Local development:

• OBJECTVERSE_VISION_BACKEND=mock
• OBJECTVERSE_TEXT_BACKEND=mock

Live Space:

• OBJECTVERSE_VISION_BACKEND=minicpm-v

• VISION_MODEL_ID=openbmb/MiniCPM-V-2_6

• OBJECTVERSE_TEXT_BACKEND=mock

• hardware: zero-a10g

• No commercial model API is used.

• GGUF files, tokens, credentials, and private images should not be committed.

Next Recommended Gate

For live vision, keep the Space on ZeroGPU with MiniCPM-V and rerun hosted validation after dependency, Space variable, or deploy changes.

Next text model gate:

Download or mount the published GGUF on the target runtime, set:

OBJECTVERSE_TEXT_BACKEND=llama-cpp
TEXT_MODEL_PATH=/absolute/path/to/objectverse-diary-qwen15b-lora-v2-q4_k_m.gguf

Then rerun the local or Space smoke path before claiming live text runtime.

Live VLM validation gate:

.venv/bin/python -B scripts/check_space_vlm.py \
  --configure-space \
  --hardware zero-a10g \
  --space-url https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary \
  --output docs/SPACE_VLM_REPORT.md \
  --json-output docs/SPACE_VLM_REPORT.json \
  --trace-output-dir data/traces/space-vlm \
  --timeout-seconds 1200

If rollback is needed, use:

.venv/bin/python -B scripts/check_space_vlm.py \
  --space-url https://huggingface.co/spaces/build-small-hackathon/ObjectverseDiary \
  --configure-space \
  --hardware zero-a10g \
  --skip-validation \
  --rollback-to-mock