docs/03-dev-schedule.md

# Objectverse Diary — 开发计划（Day-by-Day）

```text
周期：June 5 - June 15, 2026（共 11 天）
目标：完成 MVP、打磨 UI、冲全部徽章、提交视频与社交文案
```

---

## Day 1：立项 + 项目骨架

**目标：确定项目不可变范围。**

- [x] 配置 GitHub origin
- [ ] 确认并同步 GitHub repo
- [x] 创建 Hugging Face Space
- [x] 创建基础 Gradio app
- [x] 写 README 草稿
- [x] 确定英文主界面文案
- [x] 建立 `AGENTS.md`
- [x] 建立 `.codex/skills/`

---

## Day 2：MVP 交互闭环

**目标：先不管模型，跑通产品流程。**

- [x] 图片上传
- [x] 文本描述输入
- [x] personality mode 选择
- [x] mock object JSON
- [x] mock diary 输出
- [x] trace JSON 保存
- [x] share card HTML 预览
- [x] mock example gallery
- [x] MVP smoke tests
- [x] six public mock sample traces
- [x] local initial-stage acceptance script
- [x] local initial-stage completion report

交付：`Upload → Generate → Diary → Share Card → Trace`

---

## Day 3：接入 VLM

**目标：让 AI 真正看图。**

- [x] 接入 MiniCPM-V 或轻量 VLM
- [x] 输出 object understanding JSON
- [x] 做 JSON repair
- [x] 加 example gallery
- [x] 新增 Space VLM 验证脚本
- [x] 新增 ZeroGPU 兼容装饰器
- [x] ZeroGPU CUDA probe
- [ ] 缓存示例输出
- [ ] Space 真实图片验证（L4 因 HF `402 Payment Required` 阻塞；ZeroGPU CUDA probe 成功；2026-06-08 full validation reached the app but fell back to mock vision for mug/keyboard/shoe）

验收：上传杯子/键盘/鞋子，模型能识别物品并提取外观特征。

完成记录：MiniCPM-V 2.6 已作为可配置 vision backend 接入，默认仍是 mock vision；`scripts/check_space_vlm.py` 已可用三张临时公开图片验证 Space 端 mug/keyboard/shoe。2026-06-06 已尝试切到 L4，但 Hugging Face 返回 `402 Payment Required`；随后 ZeroGPU CUDA probe 成功。2026-06-08 full validation reached the app through the direct `hf.space` path, but all three objects included `vision-fallback-to-mock`。文本生成已接入可选 llama.cpp runtime wiring，但最终 GGUF 模型仍未选择/下载。

---

## Day 4：文本模型 + llama.cpp

**目标：让核心人格生成走小模型本地推理。**

- [ ] 下载小模型 GGUF
- [x] 接入可选 llama.cpp / llama-cpp-python runtime wiring
- [x] 封装 `generate_persona()`
- [x] 封装 `generate_diary()`
- [x] README 说明运行方式
- [ ] 用真实 GGUF 做本地 smoke test
- [ ] README 说明最终模型参数量

交付：`src/models/llama_cpp_runner.py` 已支持 `TEXT_MODEL_PATH`；不提交 `models/text_model.gguf`。后续仍需确定真实 GGUF、参数量和训练/发布路径。

---

## Day 5：训练数据 + 微调准备

**目标：冲 Well-Tuned 勋章。**

- [x] 设计 SFT schema
- [x] 生成 mock SFT preview 数据
- [ ] 生成 200-500 条 real/candidate object-persona 样本
- [ ] 手工精选 50 条高质量样本
- [ ] 上传 dataset 到 HF
- [ ] 准备 LoRA 训练脚本

数据格式示例：

```json
{
  "instruction": "Create a secret diary persona for this object.",
  "input": {
    "object": "old keyboard",
    "features": ["dusty", "mechanical keys", "developer desk"],
    "mode": "cynical"
  },
  "output": {
    "character_name": "Clackwell",
    "diary": "He calls it productivity. I call it percussion with anxiety.",
    "tags": ["burnout instrument", "debug witness", "plastic philosopher"]
  }
}
```

---

## Day 6：LoRA 微调 + Hub 发布

**目标：拿到可展示的自微调模型。**

- [ ] 用 Modal credits 进行训练
- [ ] 导出 LoRA adapter
- [ ] 发布 HF model repo
- [ ] app 中加入模型说明
- [ ] README 加 `Well-Tuned` section

交付：HF model repo、HF dataset repo、train log、model card

> ⚠️ Modal credits 兑换码不应公开分享，项目文档里只写"used Modal credits"。

---

## Day 7：UI 魔改

**目标：冲 Off-Brand 勋章。**

视觉方向：

```text
A strange archive room for everyday objects.
Dark paper texture, amber highlights, typewriter output, museum labels.
```

界面布局：

```text
Left:   Object Intake
Middle: Object File
Right:  Secret Diary
Bottom: Share Card + Trace
```

- [x] 自定义 CSS
- [x] 自定义 hero section
- [x] 隐藏 Gradio 默认风格
- [x] 加 typewriter / archive reveal 视觉感
- [x] 做英文主文案 + 中文辅助
- [x] 做 6 个示例卡片

完成记录：Phase 2 UI 已完成为 archive dashboard。MiniCPM-V 2.6 vision backend 和可选 llama.cpp runtime wiring 已接入但默认仍 mock；LoRA 未接入；`UI 参考/` 仅作为本地视觉参考，不入库。

---

## Day 8：Trace + Sharing is Caring

**目标：公开可复现材料。**

- [x] trace logger
- [x] sample traces
- [x] prompt templates
- [x] dataset preview
- [x] trace JSONL export
- [x] 失败案例记录
- [x] Space VLM validation report 模板
- [ ] 真实模型 traces
- [ ] GitHub repo 同步整理

---

## Day 9：Field Notes

**目标：完成技术博客。**

英文标题：`Building Objectverse Diary: A Small-Model AI Toy Where Everyday Objects Come Alive`

博客结构：

1. Why I built it
2. Why Track 2
3. Why small models are enough
4. Product design
5. Model architecture
6. Gradio Off-Brand UI
7. llama.cpp runtime
8. Fine-tuning dataset
9. Traces and reproducibility
10. What failed
11. What I would improve next

---

## Day 10：Demo 视频

**目标：视频必须比代码更能打。**

建议长度：90 秒

```text
 0- 8s  What if every object around you had a secret life?
 8-20s  This is Objectverse Diary, a small-model AI toy built with Gradio.
20-35s  Upload a photo of any everyday object.
35-50s  A vision model reads the object, then a small fine-tuned model creates its hidden personality.
50-70s  Now this coffee mug writes its secret diary and complains about its owner.
70-82s  You can chat with the object and generate a shareable personality card.
82-90s  Built with small models, Gradio, llama.cpp, public traces, and no commercial cloud APIs.
```

---

## Day 11：提交检查

- [ ] Space under official org
- [ ] Space MiniCPM-V validation passes for mug, keyboard, and shoe（当前 wired but hosted validation falls back to mock）
- [ ] Demo video ready
- [ ] Social post ready
- [ ] README complete
- [ ] Model parameter count documented
- [ ] No commercial API
- [ ] Fine-tuned model linked
- [ ] Dataset linked
- [ ] Traces linked
- [ ] Field Notes linked
- [ ] UI English-first, Chinese-second
- [ ] Submit before June 15, 2026