Add OpenSeek submission for LongContext-ICL-Annotation

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/README.md

@@ -0,0 +1,41 @@
+# OpenSeek Submission
+
+This directory contains the final open-source submission for the
+`LongContext-ICL-Annotation` track.
+
+## Included Files
+
+- `技术报告-OpenSeek.pdf`
+  - Final technical report PDF used for the competition submission.
+- `submission.zip`
+  - Exact 8-task prediction archive submitted on the platform.
+- `源代码-OpenSeek.zip`
+  - Exact source-code archive included in the final competition submission.
+- `code/`
+  - Extracted source tree for direct review, including configs, prompts,
+    scripts, tests, and usage documentation.
+
+## Method Summary
+
+The solution uses `Qwen3-4B` as the only large language model and uses
+`FlagScale` as the required runtime framework. The pipeline combines:
+
+- long-context example retrieval and compression
+- front-back evidence reordering
+- multi-protocol first-pass inference
+- confidence-based recheck and adjudication
+- deterministic post-processing for selected tasks
+- submission validation and packaging
+
+The final confirmed platform score for this submission line is `87.35`.
+
+## Reproduction
+
+See [code/README.md](code/README.md) for environment setup, data preparation,
+FlagScale deployment, inference, evaluation, packaging, and validation.
+
+## Scope
+
+This submission only adds files under:
+
+`openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/`

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/.gitignore

@@ -0,0 +1,38 @@
+# Python and local tooling
+.venv/
+venv/
+__pycache__/
+*.py[cod]
+*.pyo
+*.pyd
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+
+# OS/editor noise
+.DS_Store
+Thumbs.db
+*.swp
+*.swo
+
+# OpenBayes workspace-local state
+.openbayesgear
+
+# Official data and generated caches.
+# Keep only README/.gitkeep placeholders under version control.
+data/raw/*
+!data/raw/.gitkeep
+data/processed/*
+!data/processed/.gitkeep
+data/openbayes_cache/*
+
+# Runtime outputs, logs, pulled remote artifacts, reports, submissions, and zips.
+# Release metadata that should be versioned belongs under docs/.
+outputs/*
+submission*.zip
+*.zip
+
+# Generated document intermediates
+report/.technical_report.normalized-headings.md

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/.openbayesignore

@@ -0,0 +1,20 @@
+.git
+.git/
+.git/**
+.pytest_cache
+.pytest_cache/
+__pycache__
+__pycache__/
+*.pyc
+.venv
+.venv/
+outputs
+outputs/
+outputs/**
+data/raw
+data/raw/
+data/raw/**
+data/processed
+data/processed/
+data/processed/**
+*.zip

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/CODE_PACKAGE_README.md

@@ -0,0 +1,17 @@
+# Code Package Notes
+
+This archive is the official code package for submission review.
+
+Included:
+- source code under `src/`
+- runnable scripts under `scripts/`
+- configs under `configs/`
+- prompts under `prompts/`
+- tests under `tests/`
+- `README.md` and `requirements.txt`
+
+Intentionally excluded:
+- generated outputs and caches
+- temporary folders and local review artifacts
+- report sources and PDF exports
+- internal release notes and development-only documents

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/Makefile

@@ -0,0 +1,25 @@
+PYTHON ?= python3
+CONFIG ?= configs/base.yaml
+
+.PHONY: prepare predict eval package test release-check release-check-dirty
+
+prepare:
+	./scripts/import_official_data.sh $(DATA_DIR)
+
+predict:
+	$(PYTHON) -m src.main predict --config $(CONFIG)
+
+eval:
+	$(PYTHON) -m src.main evaluate --config $(CONFIG)
+
+package:
+	$(PYTHON) -m src.main package --config $(CONFIG)
+
+test:
+	$(PYTHON) -m pytest -q
+
+release-check:
+	$(PYTHON) scripts/release_preflight.py
+
+release-check-dirty:
+	$(PYTHON) scripts/release_preflight.py --allow-dirty

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/README.md

@@ -0,0 +1,125 @@
+# AI-LAB OpenSeek Submission
+
+面向 OpenSeek `LongContext-ICL-Annotation` 赛题的可复现参赛工程。
+
+本项目基于 `Qwen3-4B` 构建统一的 8 任务自动标注流程，核心链路包括：
+
+- 长上下文示例检索与压缩
+- `front-back` 证据重排
+- 多协议首轮推理
+- 一致性与置信度判断
+- 低置信样本复核
+- 提交校验与统一打包
+
+## 环境要求
+
+- Python 3.10+
+- 可用的 Qwen3-4B 权重目录
+- 官方比赛数据目录
+- 若采用正式部署路径，使用 FlagScale 作为模型加载与服务框架
+
+安装依赖：
+
+```bash
+python3 -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+```
+
+## 目录说明
+
+```text
+AI-LAB/
+├── configs/         运行配置
+├── data/            数据目录约定与缓存目录
+├── prompts/         各任务提示词协议
+├── scripts/         运行、校验与打包脚本
+├── src/             主流程实现
+├── tests/           基础回归测试
+└── requirements.txt 环境依赖
+```
+
+## 数据准备
+
+将官方数据导入到 `data/raw/openseek/`：
+
+```bash
+./scripts/import_official_data.sh /path/to/LongContext-ICL-Annotation/data
+```
+
+## 快速运行
+
+本地 smoke：
+
+```bash
+./scripts/run_infer.sh configs/local_smoke.yaml
+./scripts/run_eval.sh configs/local_smoke.yaml
+./scripts/package_submission.sh configs/local_smoke.yaml
+```
+
+标准运行：
+
+```bash
+./scripts/run_infer.sh configs/base.yaml
+./scripts/run_eval.sh configs/base.yaml
+./scripts/package_submission.sh configs/base.yaml
+```
+
+## FlagScale 部署路径
+
+若采用 FlagScale 服务化推理，先按实际模型路径修改
+`configs/flagscale_serve.template.yaml`，然后启动服务：
+
+```bash
+flagscale run -p configs -n flagscale_serve.template -a run \
+  serve.0.engine_args.model=/path/to/Qwen3-4B \
+  serve.0.engine_args.port=2026 \
+  +serve.0.engine_args.max_model_len=32768 \
+  +serve.0.engine_args.max_num_seqs=1 \
+  serve.0.engine_args.gpu_memory_utilization=0.88
+```
+
+随后运行正式配置：
+
+```bash
+./scripts/run_infer.sh configs/openbayes_flagscale_full.yaml
+./scripts/run_eval.sh configs/openbayes_flagscale_full.yaml
+./scripts/package_submission.sh configs/openbayes_flagscale_full.yaml
+```
+
+## 结果校验
+
+任务 8 静态检查：
+
+```bash
+python3 scripts/check_task8_predictions.py outputs/predictions/openseek-8-v1.jsonl
+```
+
+任务 8 运行代理检查：
+
+```bash
+python3 scripts/check_task8_runtime.py outputs/predictions/openseek-8-v1.jsonl
+```
+
+完整发布预检：
+
+```bash
+make release-check
+```
+
+## 关键文件
+
+- 主流程：`src/ai_lab/pipeline.py`
+- 官方数据读取：`src/ai_lab/adapters/official_reader.py`
+- 检索与重排：`src/ai_lab/retrieval/`
+- 决策与投票：`src/ai_lab/decision/`
+- 输出解析：`src/ai_lab/output_parser.py`
+- 提交校验：`src/ai_lab/submit/validate_submission.py`
+
+## 当前交付物
+
+在本次 OpenSeek 官方仓库提交中，对应交付物位于上一级目录：
+
+- 最终预测包：`../submission.zip`
+- 最终代码包：`../源代码-OpenSeek.zip`
+- 正式技术报告 PDF：`../技术报告-OpenSeek.pdf`

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/configs/base.yaml

@@ -0,0 +1,78 @@
+experiment:
+  team_name: your-team
+  project_name: longcontext-icl-annotation
+  seed: 42
+
+model:
+  name: Qwen/Qwen3-4B
+  backend: heuristic
+  device: cuda
+  dtype: bfloat16
+  load_in_4bit: true
+  bnb_4bit_compute_dtype: float16
+  bnb_4bit_quant_type: nf4
+  bnb_4bit_use_double_quant: true
+  max_context_tokens: 32768
+  max_new_tokens: 512
+  temperature: 0.0
+  top_p: 0.95
+  repetition_penalty: 1.0
+  trust_remote_code: true
+  model_path: models/Qwen3-4B
+  tokenizer_path: models/Qwen3-4B
+  api_url: http://0.0.0.0:2026/v1/completions
+  api_model_name: Qwen3-4B
+  timeout_seconds: 600
+
+data:
+  registry_path: configs/datasets/registry.yaml
+  raw_root: data/raw
+  official_data_dir: data/raw/openseek
+  processed_root: data/processed
+  prediction_root: outputs/predictions
+
+icl:
+  selector: lexical_topk
+  num_examples: 64
+  chunk_budget_tokens: 24000
+  reserve_generation_tokens: 1024
+  multi_turn_memory: summary_trace
+  example_template: plain
+  prefer_long_context: true
+  small_text_threshold_chars: 1800
+  chunk_size_chars: 896
+  chunk_overlap_chars: 96
+  retrieval_top_k: 8
+
+prompt:
+  system_prompt_path: configs/prompts/system_prompt.txt
+  output_format_path: configs/prompts/output_format.txt
+  protocol_a_path: prompts/protocol_a.yaml
+  protocol_b_path: prompts/protocol_b.yaml
+  protocol_c_light_path: prompts/protocol_c_light.yaml
+  protocol_c_path: prompts/protocol_c.yaml
+  protocol_a_variant_path: prompts/variants/protocol_a_v2.yaml
+  code_generation_protocol_path: prompts/code_generation.yaml
+
+decision:
+  accept_threshold: 0.82
+  variant_threshold: 0.82
+  adjudication_threshold: 0.68
+
+runtime:
+  log_dir: outputs/logs
+  overwrite: true
+  save_prompts: false
+  max_samples_per_task: null
+
+submission:
+  zip_name: submission.zip
+  required_prediction_files:
+    - openseek-1-v1.jsonl
+    - openseek-2-v1.jsonl
+    - openseek-3-v1.jsonl
+    - openseek-4-v1.jsonl
+    - openseek-5-v1.jsonl
+    - openseek-6-v1.jsonl
+    - openseek-7-v1.jsonl
+    - openseek-8-v1.jsonl

openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/configs/datasets/registry.yaml

@@ -0,0 +1,41 @@
+datasets:
+  - task_id: 1
+    file_name: openseek-1_closest_integers.json
+    task_name: closest_integers
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 2
+    file_name: openseek-2_count_nouns_verbs.json
+    task_name: count_nouns_verbs
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 3
+    file_name: openseek-3_collatz_conjecture.json
+    task_name: collatz_conjecture
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 4
+    file_name: openseek-4_conala_concat_strings.json
+    task_name: conala_concat_strings
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 5
+    file_name: openseek-5_semeval_2018_task1_tweet_sadness_detection.json
+    task_name: semeval_2018_task1_tweet_sadness_detection
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 6
+    file_name: openseek-6_mnli_same_genre_classification.json
+    task_name: mnli_same_genre_classification
+    task_type: classification
+    min_context_length: 30000
+  - task_id: 7
+    file_name: openseek-7_jeopardy_answer_generation_all.json
+    task_name: jeopardy_answer_generation_all
+    task_type: generation
+    min_context_length: 30000
+  - task_id: 8
+    file_name: openseek-8_kernel_generation.json
+    task_name: kernel_generation
+    task_type: code_generation
+    min_context_length: 16000