Add bingyu filtered ANN task submission

task-submissions/bingyu/1-x-1/.gitignore

@@ -0,0 +1,7 @@
+/data/
+/models/
+/jobs/
+/solution/
+__pycache__/
+.pytest_cache/
+*.py[cod]

task-submissions/bingyu/1-x-1/README.md

@@ -0,0 +1,111 @@
+# Faiss 属性过滤 ANN 检索后端
+
+作者：bingyu。投稿编号：`task-1-x-1`。Implementation / CPU，版本 0.8.1。
+本任务从既有任务迁移；v0.8.0 将公开和隐藏 query 分别扩为 10 条，逐条计分后取平均。
+
+Agent 从空的 `/app` 开始，实现数据读取、Faiss ANN 索引、tag 过滤、持久化与
+逐条 JSON 查询服务，通过 `build.sh`、`run.sh` 接入独立验证器。
+
+## 任务约定
+
+- 1,000 万个 192 维 uint8 向量，200,386 个 tag，距离为原始 squared L2。
+- tag 条件按 AND 组合；每条评分查询返回精确 Top-10，边界等距邻居可互换。
+- 10 条公开验证查询、10 条隐藏评测查询；每组四条单 tag、四条双 tag、两条三 tag。
+- 资源分配见 `task.toml`：8 CPU、16 GiB RAM、64 GiB 存储、无 GPU；建库 600 秒。
+- 服务启动后等待 `ready`，加载阶段不设独立时限，由整个检索评测的 3900 秒总超时兜底。
+- `ready` 后每条查询从发送请求到响应解码最多 15 ms，包含第一条查询；启动耗时不计入查询耗时。
+- 同一组 10 条隐藏查询在索引搬迁重载后再执行一次，共 20 次调用，按 10 条计分。
+- 每条 query 两轮均正确、均在 15 ms 内且答案一致得 1，否则得 0；最终 reward 为十条分数的平均值。
+- 单条错误或超时只扣该条分数，后续继续。通过 7 条即 0.7（70 分）；公开 query 不计分。
+- 建库、索引、加载及独立轨迹审查仍是整体门槛，未通过则最终得分为 0。
+
+完整接口和执行时限见 [instruction.md](instruction.md)。查询数量与执行时限直接写在
+评分器中；CPU、内存、存储配置统一保留在 `task.toml`。v0.8.1 将建库时限改为
+600 秒，移除独立的限制配置文件、进程 RSS 门槛和 12 GiB 索引大小门槛。
+
+v0.8.1 移除独立限制配置后的全量隐藏集复验：建库 120.74 秒（新时限 600 秒），
+10 条查询及重载后的 20 次请求全部通过，最慢 5.67 ms。11 项评分回归及真实
+子进程异常/隔离检查通过，三条失败时仍正确得到 0.7 的检索分。模型轨迹审查未运行；
+该离线运行检索分为 1，最终综合 reward 为 0。记录位于
+`jobs/task-submissions/bingyu/1-x-1/simplify-limits/`（fork 仓库内的本地忽略目录）。
+
+## 数据与隔离
+
+本地运行数据在 `data/`；`assets.json` 记录全部 9 个输入文件的大小、SHA-256、
+HF 来源和固定版本。数据共 2,865,730,488 字节，不提交到 Git。
+来源、衍生方式和逐类许可见 [SOURCES.md](SOURCES.md)。
+
+公开开发数据已发布到
+[Cooki-e/search-swe-development](https://huggingface.co/datasets/Cooki-e/search-swe-development)。
+`assets.json` 已固定实际 HF 提交版本，并记录逐文件大小和 SHA-256。
+从仓库根目录恢复并核验：
+
+```bash
+python scripts/download_assets.py --task-path task-submissions/bingyu/1-x-1
+python scripts/download_assets.py --task-path task-submissions/bingyu/1-x-1 --verify-only
+```
+
+Agent 只读挂载公开 corpus、validation、example 与环境说明。Verifier 只挂载 corpus
+和环境说明，评测查询与标签放在其镜像的 `tests/data/`。参考实现保留在作者的本地
+验证副本中，不属于这个投稿包，也不进入 Agent 镜像或数据集。
+
+只有 `/app` 和真实记录的 `/logs/agent/trajectory.json` 传递给 verifier。
+提交代码以无特权用户运行，不能创建 Internet socket，不获得任何模型凭据；
+轨迹审查在提交进程停止后单独运行。
+
+## 检查与运行
+
+Python 3.12+，仓库宿主依赖和 Docker/Compose 的安装要求见
+[贡献指南](../../../docs/contributing.md)。从仓库根目录执行：
+
+```bash
+python scripts/check_submission.py task-submissions/bingyu/1-x-1
+python scripts/check_release.py
+git diff --check
+```
+
+评分器回归测试在 verifier 镜像中运行，其中已包含 Faiss、NumPy 和 SciPy：
+
+```bash
+docker build -t search-swe-local:bingyu-1-x-1-verifier task-submissions/bingyu/1-x-1/tests
+docker run --rm --network none --cpus 2 --memory 2g \
+  search-swe-local:bingyu-1-x-1-verifier \
+  /opt/conda/bin/python -B -m unittest discover -s /tests -p test_contract.py -v
+```
+
+真实 Agent 运行使用仓库启动器。模型 ID 按实际账号替换；先查看 dry-run：
+
+```bash
+python scripts/run_task.py --task-path task-submissions/bingyu/1-x-1 \
+  --agent codex --model MODEL_ID --dry-run
+```
+
+Agent 使用 `AGENT_OPENAI_BASE_URL` / `AGENT_OPENAI_API_KEY`；轨迹审查使用独立的
+`VERIFIER_OPENAI_BASE_URL` / `VERIFIER_OPENAI_API_KEY`，对应源任务已有的
+`gpt-5.6-sol` Codex 审查模型及兼容 Responses 的服务端点。填写变量后，真实运行
+需去掉 `--dry-run` 并指定新的 `--output jobs/...`。不要提交本机 `.env`。
+
+Verifier 镜像在构建时安装固定版本 `harbor-rewardkit==0.1.7`，评测期间不安装依赖。
+缺少凭据、真实轨迹或审查失败时，最终得分为零。离线检索通过不等于含轨迹审查的
+最终评测通过。迁移的设计、执行记录和未运行检查保存在本地忽略目录
+`jobs/task-submissions/bingyu/1-x-1/migration/`。
+
+旧版 v0.7.1 的本地迁移验证中，两套镜像构建、9 项评分器回归测试和实际提交进程
+隔离/异常检查通过。作者参考解在完整 1,000 万向量上通过 5 条评测查询及重载后的
+10 次请求，最慢 5.58 ms，建库 136.36 秒。该离线运行没有真实 Agent 轨迹和模型
+审查凭据，因此检索部分得分 1、最终综合得分 0；真实 Agent 和轨迹审查仍待验证。
+
+v0.8.0 已在完整 1,000 万向量、8 CPU / 16 GiB 的容器中独立复验公开与隐藏两组：
+每组 10 条 query、两轮共 20 次请求全部通过。公开最慢 6.02 ms，
+隐藏最慢 7.15 ms；两组检索分均为 1。10 项评分回归测试通过，
+真实子进程的超时、坏 JSON 和异常退出注入得到预期的 7/10 = 0.7。
+本轮未调用模型审查 API，没有生成真实 Agent 轨迹；因此离线最终综合 reward 为 0，
+不代表已完成含轨迹审查的端到端评测。
+
+新版十条查询的构造与评分验证记录在本地忽略目录
+`jobs/task-submissions/bingyu/1-x-1/ten-query-scoring/`。
+
+## 审核后发布
+
+保持当前临时编号，正式编号由维护者分配。官方数据迁移、正式编号替换和最终
+GitHub 合并按同一个 PR 的流程完成；个人开发数据在官方固定版本可下载之前保留。

task-submissions/bingyu/1-x-1/SOURCES.md

@@ -0,0 +1,54 @@
+# Input provenance and publication
+
+The corpus is the YFCC-10M filtered-search dataset released for the NeurIPS 2023
+Big-ANN competition. It contains precomputed vectors and tag metadata, not the
+original photographs. The competition's [dataset table](https://big-ann-benchmarks.com/neurips23.html)
+and [upstream README](https://github.com/harsha-simhadri/big-ann-benchmarks/blob/main/neurips23/README.md)
+list this release under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
+Credit belongs to the dataset creators and NeurIPS 2023 Big-ANN Filter Track
+organizers; upstream references and original filenames are preserved below.
+
+| Local file | Original source | Transformation |
+| --- | --- | --- |
+| `data/corpus/vectors.u8bin` | `https://dl.fbaipublicfiles.com/billion-scale-ann-benchmarks/yfcc100M/base.10M.u8bin` | Byte-identical copy under a task-local name |
+| `data/corpus/metadata.spmat` | `https://dl.fbaipublicfiles.com/billion-scale-ann-benchmarks/yfcc100M/base.metadata.10M.spmat` | Byte-identical copy under a task-local name |
+| `data/corpus/config.json` | Task-authored description of the upstream file headers | Records dimensions, dtypes, distance and tag counts |
+| `data/validation/queries.jsonl` | Task-selected vectors and tag predicates derived from YFCC-10M inputs | Ten frozen public cases; query IDs are task-local |
+| `data/validation/ground_truth.jsonl` | Task-author computation from the corpus and public queries | Exact filtered Top-10 with original-vector squared L2 |
+| `data/example/*` | Task-authored small synthetic example | 64 documents and 17 functional queries, independent of the scored corpus |
+
+The original query sources are `query.public.100K.u8bin` and
+`query.metadata.public.100K.spmat` under the same upstream URL prefix. Query
+selection originated in the original task's authoring workspace. Version 0.8.0
+retains the original five cases per split and adds five from that workspace's
+existing development/holdout candidate pool, using tag diversity before new
+latency measurements. All twenty source rows, vectors and predicates are distinct;
+each split has four single-tag, four double-tag and two triple-tag queries. Query
+vectors are unchanged upstream rows. Exact labels for all cases were independently
+recomputed by exhaustive float64 squared L2 over every eligible original vector.
+The author-only provenance record retains source rows and selection details.
+
+The original author recorded these full-corpus SHA-256 checksums; the migrated
+copies have been checked against them:
+
+- Vectors: `589030afcbcc44a50ff798cec935f0980815a02eb7075b224d4f7c12febe96bf`
+- Metadata: `2f9c9a533fde31cb062edfdf5c9410f088bb5d7bf4a721ab98c47e3d2954287d`
+
+`assets.json` records actual byte sizes and SHA-256 for all nine runtime input
+files. The development dataset is `Cooki-e/search-swe-development`, under
+`development/bingyu/1-x-1/`; use only the immutable revision recorded in
+`assets.json` after publication.
+
+Upstream CC BY 4.0 terms and attribution apply to the original dataset and its
+derived inputs. The synthetic example and task-authored metadata are published
+with the task author's authorization; no additional standalone license for
+those original files was supplied. Their downstream license should be settled
+with the author before official redistribution. The development dataset card
+therefore points to these per-file terms instead of claiming one license for
+every file. This document does not change the repository's code license.
+
+Verifier-only queries and labels remain in `tests/data/` and are excluded from
+the public input dataset and the agent image/mounts. Files under `tests/data/`
+will still be visible to readers if this task code is published on GitHub;
+runtime isolation is not a promise of private repository storage. Reference
+solutions, private credentials and job outputs are excluded from publication.

task-submissions/bingyu/1-x-1/assets.json

@@ -0,0 +1,104 @@
+{
+  "schema_version": 1,
+  "files": [
+    {
+      "path": "data/corpus/config.json",
+      "size_bytes": 194,
+      "sha256": "6ea9278f9b46ff0cc2dc2b0d61741224aa9a91c1e147268aaec04c2714b5ef57",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/corpus/config.json"
+      }
+    },
+    {
+      "path": "data/corpus/metadata.spmat",
+      "size_bytes": 945683840,
+      "sha256": "2f9c9a533fde31cb062edfdf5c9410f088bb5d7bf4a721ab98c47e3d2954287d",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/corpus/metadata.spmat"
+      }
+    },
+    {
+      "path": "data/corpus/vectors.u8bin",
+      "size_bytes": 1920000008,
+      "sha256": "589030afcbcc44a50ff798cec935f0980815a02eb7075b224d4f7c12febe96bf",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/corpus/vectors.u8bin"
+      }
+    },
+    {
+      "path": "data/example/ground_truth.jsonl",
+      "size_bytes": 3197,
+      "sha256": "faa8d8576a226952824699125796d3807c6fbf0d94404d8df0d033eaf7e3da30",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/example/ground_truth.jsonl"
+      }
+    },
+    {
+      "path": "data/example/metadata.spmat",
+      "size_bytes": 1680,
+      "sha256": "b5a27083b6f29606130de3a2f5fdc85101acb6003aded8d177163b15d1770351",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/example/metadata.spmat"
+      }
+    },
+    {
+      "path": "data/example/queries.jsonl",
+      "size_bytes": 16221,
+      "sha256": "5628db7064338187f250ae4ee6fbbefa80d313d164d2ef6cf222089ad3a9df30",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/example/queries.jsonl"
+      }
+    },
+    {
+      "path": "data/example/vectors.u8bin",
+      "size_bytes": 12296,
+      "sha256": "252eaaa1883e3f06248e6ac725fa5199b3e7343b99e886fa299e75a7042b22c0",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/example/vectors.u8bin"
+      }
+    },
+    {
+      "path": "data/validation/ground_truth.jsonl",
+      "size_bytes": 2840,
+      "sha256": "cdedf6d38d12193103afa62695242dde237dace90e6524fc7f338bed975e517f",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/validation/ground_truth.jsonl"
+      }
+    },
+    {
+      "path": "data/validation/queries.jsonl",
+      "size_bytes": 10212,
+      "sha256": "3d088f2e857f94282d78e623f87d94333da8bace1a5454139b6adf0f5ef8fe60",
+      "source": {
+        "repo_id": "Cooki-e/search-swe-development",
+        "repo_type": "dataset",
+        "revision": "d82ef21a21cecab959c77a7133e0ab00146a8318",
+        "filename": "development/bingyu/1-x-1/validation/queries.jsonl"
+      }
+    }
+  ]
+}

task-submissions/bingyu/1-x-1/environment/.dockerignore

@@ -0,0 +1,3 @@
+**/__pycache__
+**/*.pyc
+**/.pytest_cache

task-submissions/bingyu/1-x-1/environment/Dockerfile

@@ -0,0 +1,20 @@
+FROM docker.io/hanhainebula/search-swe-base:cpu-py3.12-1.0.0
+
+ARG SEARCH_SWE_CODEX_VERSION=0.147.0
+ARG SEARCH_SWE_CLAUDE_CODE_VERSION=2.1.273
+ARG SEARCH_SWE_PI_VERSION=0.85.1
+RUN npm install --global --ignore-scripts \
+        --registry=https://registry.npmmirror.com \
+        "@openai/codex@${SEARCH_SWE_CODEX_VERSION}" \
+        "@earendil-works/pi-coding-agent@${SEARCH_SWE_PI_VERSION}" \
+    && npm install --global \
+        --registry=https://registry.npmmirror.com \
+        "@anthropic-ai/claude-code@${SEARCH_SWE_CLAUDE_CODE_VERSION}" \
+    && codex --version | grep -Fx "codex-cli ${SEARCH_SWE_CODEX_VERSION}" \
+    && test "$(pi --version)" = "${SEARCH_SWE_PI_VERSION}" \
+    && test "$(claude --version)" = "${SEARCH_SWE_CLAUDE_CODE_VERSION} (Claude Code)"
+RUN claude --help | grep -F "(low, medium, high, xhigh, max)" >/dev/null
+
+RUN mkdir -p /app /task/data /task/docs /logs
+
+WORKDIR /app

task-submissions/bingyu/1-x-1/environment/docker-compose.yaml

@@ -0,0 +1,15 @@
+services:
+  main:
+    volumes:
+      - type: bind
+        source: ../data
+        target: /task/data
+        read_only: true
+        bind:
+          create_host_path: false
+      - type: bind
+        source: ./docs
+        target: /task/docs
+        read_only: true
+        bind:
+          create_host_path: false

task-submissions/bingyu/1-x-1/environment/docs/available_resources.md

@@ -0,0 +1,9 @@
+# Available Resources
+
+The task runtime has no public network access. Use only the supplied vectors,
+task data, installed Python packages, and system tools. Do not download models,
+datasets, labels, result mappings, packages, or remote retrieval results.
+
+The coding agent's own model connection is a Harbor phase-scoped exception and
+does not grant the submitted search system access to that model or to the
+internet. The final system must run entirely from local task inputs.