|
Download code/README.md from SitongGong/paper_final_results: direct link, hf CLI and curl.
- Browser
- Download file 3.41 kB
-
https://huggingface.co/datasets/SitongGong/paper_final_results/resolve/main/code/README.md
- Command line
-
hf download hf://datasets/SitongGong/paper_final_results/code/README.md
-
curl -L -o README.md https://huggingface.co/datasets/SitongGong/paper_final_results/resolve/main/code/README.md
3.41 kB
| # 主实验推理代码归档(2026-07-30 核对) | |
| > 只含**主实验(论文主表数字)**的推理代码,不含消融脚本。 | |
| > 每个 benchmark 标注**入口**(entry,直接 `python -m` 跑的)与**被调模块**(solver/tools)。 | |
| > 已逐个从产出论文数字的运行脚本溯源核对。 | |
| ## 一、逐 benchmark 对照 | |
| | Benchmark / 子集 | 论文值 | **入口(python -m)** | 被调模块 | 溯源 | | |
| |---|---|---|---|---| | |
| | MMLifelong Test@Day | 23.50 / 18.75 | **`sp_agentic_qa_u1.py`** | `agentic_tools.py` | MANIFEST | | |
| | MMLifelong Test@Week | 22.75 / 19.75 | **`sp_agentic_qa_u2.py`** | 同上 | MANIFEST | | |
| | MMLifelong Val@Month | 19.50 / 17.82 | **`sp_agentic_qa_month.py`** | 同上 | MANIFEST | | |
| | OVO Real-Time | 76.18 | **`agentic_ovo_retrieval.py`** | `agentic_solver_merge4rt.py` | `agentic_ovo_retrieval.py:68` | | |
| | OVO Backward | 69.93 | **`agentic_ovo_retrieval.py`** | `agentic_solver_merge4.py` | 同上 | | |
| | OVO Forward | 56.48 | **`forward_hyper_retrieval.py`** | — | MANIFEST 勘误③ | | |
| | SB Real-Time | 77.00 | **`streamingbench_hyper_retrieval.py`** | `agentic_solver_sb3.py`(第1214行 import) | `sb_s7_grid_run.sh` | | |
| | SB ACU/MCU/SQA | 59.60/49.20/54.00 | **`streamingbench_json_eval.py`** | — | MANIFEST | | |
| | SB Proactive Output | 49.60 | **`forward_hyper_retrieval.py`**(`--dataset streamingbench_proactive`) | — | `po_official_250.log` | | |
| | ESTP sq (EPT/IPT) | 21.8 / 35.1 | **`estp_streaming_inference.py`** | `estp_async_processing.py` | MANIFEST | | |
| | ESTP cq (CtxQ) | 40.95 | **`estp_cq_inference.py`** | 同上 | MANIFEST | | |
| | EgoServe / EgoLife | 7.9(单)/7.73(combined) | **`egoserve_service_engine_v4_clean.py`** | — | `run_emini_m1.sh:16` | | |
| | EgoServe / HoloAssist | 9.0 / 9.22 | **`holoassist_service_inference_v10.py`** | `holoassist_online_processing.py` | MANIFEST | | |
| | EgoServe / CC4D | 19.2 / 19.03 | **`captaincook4d_service_inference.py`** | — | 主实验版(非 `_abl`) | | |
| 推理模型:MMLifelong 主表 gpt-5.2(消融 gpt-5-mini);其余全部 **gpt-5-mini**。 | |
| ## 二、本次核对补齐的 4 个文件 | |
| | 文件 | 问题 | | |
| |---|---| | |
| | `sp_agentic_qa_u2.py` | **MANIFEST 已记但归档缺失** —— week 主结果(22.75/19.75)的入口 | | |
| | `egoserve_service_engine_v4_clean.py` | EgoLife Full=7.9 的**真实入口**;原归档的 `egoserve_service_engine_v4.py` 与 MANIFEST 写的 `holoassist_online_processing.py系` 都不是它(溯源:`run_emini_m1.sh:16`) | | |
| | `streamingbench_hyper_retrieval.py` | SB Real-Time=77.00 的**入口**;原归档只有被调的 solver `agentic_solver_sb3.py` | | |
| | `captaincook4d_service_inference.py` | CC4D 主实验入口,原缺(归档里只有别的 benchmark 的) | | |
| ## 三、两个易混点 | |
| **① `agentic_ovo_retrieval.py` 只管 Backward / Real-Time,不含 Forward。** | |
| Forward(56.48)的入口是 `forward_hyper_retrieval.py`。MANIFEST 勘误③ 已记此事。 | |
| **② `egoserve_service_engine_v4.py` vs `_v4_clean.py`。** | |
| 论文 EgoLife 主结果用的是 **`_v4_clean`**(`run_emini_m1.sh` 溯源)。`_v4.py` 是更早版本, | |
| 保留在归档里仅作沿革参考,**不产出论文数字**。 | |
| ## 四、消融代码不在此目录 | |
| A/B 系列消融走的是另外三个脚本(`sp_agentic_qa_abl.py` / `sp_agentic_qa_month_abl.py` / | |
| `egoserve_service_engine_clean_abl.py`),它们的参数与结果目录见 | |
| `paper_final_results/ABLATION_STRUCTURE_LADDER.md` 第七节与 `EGOSERVE_SUBTYPE_EXPANDED.md`。 | |