Commit ef9b24f8 ef9b24f83aa45d9ac32f0f9096b6b991d706a9bb by cnb.bofCdSsphPA

Preserve proof that the cap64 benchmark has started before it finishes

Constraint: The new cap64 run is still in-flight, so only startup and stage-transition evidence can be documented safely
Rejected: Wait for cap64 results before checkpointing | Would leave the next session without a verified handoff that the larger benchmark is already running
Confidence: high
Scope-risk: narrow
Directive: Keep cap64 artifacts out of git and update strategy guidance only after report.json lands
Tested: Verified the cap64 ab_smoke process is running, confirmed the high_energy smoke-local branch entered train.py on /tmp/ab_smoke_seg_cap64_top2/high_energy/fma/manifests, and recorded the active work root and parameters in docs
Not-tested: Final cap64 metrics, hybrid branch execution, and any post-cap64 strategy conclusion
1 parent d1f13203
## 2026-06-02 cap64 benchmark 启动 checkpoint
完成项:
- 已启动新的真实 FMA `cap64` 对照 benchmark。
- 本轮配置:`subset_size=64`, `max_test_queries=32`, `seed=42`
- 当前已确认流程进入 `high_energy` 训练阶段。
验证证据:
- 主进程:`scripts/ab_smoke_segmentation.py --work-root /tmp/ab_smoke_seg_cap64_top2`
- 子流程:`external_adapters.py smoke-local ... /tmp/ab_smoke_seg_cap64_top2/high_energy`
- 当前训练:`train.py --data /tmp/ab_smoke_seg_cap64_top2/high_energy/fma/manifests ... --segment-strategy high_energy`
说明:
- 截至本 checkpoint,cap64 结果尚未生成。
- 本次提交的目的是为下一 session 固化新一轮 benchmark 已正式启动的证据。
## 2026-06-02 cap48 seed999 完结与三 seed 聚合 checkpoint
完成项:
......
......@@ -64,3 +64,5 @@ cd /workspace/acr-engine
- 已补记 `hybrid` seed=999 的中间结果:`top1=0.875 / topk=1.0 / num_queries=24`
- 已补齐 `seed=999` 最终结果,并完成 cap48 三 seed aggregate 归纳。
- 已记录 cap64 benchmark 已启动,并确认进入 `high_energy` 训练阶段。
......
......@@ -56,3 +56,9 @@ pgrep -af 'ab_smoke_seg_cap48_top2_seed999|external_adapters.py smoke-local fma
```bash
test -f /tmp/ab_smoke_seg_cap48_top2_seed999/report.json && cat /tmp/ab_smoke_seg_cap48_top2_seed999/report.json || echo NO_REPORT
```
## 下一轮已启动
- 新 benchmark:`/tmp/ab_smoke_seg_cap64_top2`
- 当前阶段:`high_energy` 训练中
- 下一 session 应优先检查 `report.json` 是否生成
......
......@@ -240,10 +240,10 @@
- `hybrid``mean_top1=0.8750, min=0.7917, max=0.9583, stdev=0.0680`
### 最优先待办
1. 基于 3-seed 结果继续设计 cap64 benchmark
2. 增加 bucket/style-aware benchmark
3. 继续优化 `hybrid`,重点降低波动并提升 hard case 稳定性
4. 提交并推送后继续下一轮验证
1. 跟进正在运行的 cap64 benchmark:`/tmp/ab_smoke_seg_cap64_top2/report.json`
2. 在 cap64 完成后更新 `open-dataset-workflow.md / session-handoff.md / CHANGELOG.md`
3. 接着增加 bucket/style-aware benchmark
4. 继续优化 `hybrid`,重点降低波动并提升 hard case 稳定性
### 续跑时不要做的事
- 不要 `git add .`
......@@ -669,3 +669,12 @@ seed123 最终结论:
- `hybrid` 上限更高但波动更大
- 最终默认策略要看更多 seed 聚合结果
## 100. 新一轮验证已启动:cap64
- 已启动:`/tmp/ab_smoke_seg_cap64_top2`
- 配置:`subset_size=64`, `max_test_queries=32`, `seed=42`
- 当前最新证据:
- 已进入 `high_energy``train.py` 阶段
- 尚未产出最终 `report.json`
......