Add LongContext-ICL-Annotation submission for team 大力水手 (Popeye)

[SweetDeathh]

Submission: 大力水手 (Popeye) — LongContext-ICL-Annotation

This PR adds the open-source submission of team 大力水手 (Popeye) for the
LongContext-ICL-Annotation track, under
openseek/competition/LongContext-ICL-Annotation/submissions/popeye/.

Files

• README.md — submission overview (English)
• 技术报告-大力水手.pdf — final technical report (Chinese)
• final_submission_v48.zip — the 8-task .jsonl prediction archive
  (openseek-1-v1.jsonl … openseek-8-v1.jsonl directly at zip root)
• code/ — full reproducible source code, data, environment files,
  deployment script, and supporting documentation

Method summary

The solution stays on the official Qwen3-4B + YaRN long-context baseline
and gains score through a controlled "main-task method line + changed-row
audit + ultra-narrow factual patch" workflow. Per-version evidence is
preserved under code/docs/phase-summaries/.

Reproduction

See code/README.md for full instructions.

Scope

Only adds files under submissions/popeye/. No existing files are modified.

[gemini-code-assist]

Code Review

This pull request adds the complete submission of team "大力水手 (Popeye)" for the LongContext-ICL-Annotation track, including reproducible source code, datasets, environment files, deployment scripts, and extensive documentation of their iterative, stability-gated method. The review feedback identifies critical issues in the setup scripts and dependency files, specifically pointing out incorrect relative path calculations in bootstrap_h20.sh and several non-existent pinned package versions (such as PyTorch, scikit-learn, regex, openai, and transformers) in bootstrap_h20.sh, requirements.txt, and environment_h20.yml that would cause installation and runtime failures.

[gemini-code-assist]

The relative path calculation for OPENSEEK_ROOT and FLAGOS_ROOT is incorrect. The script is located at openseek/competition/LongContext-ICL-Annotation/submissions/popeye/code/bootstrap_h20.sh. Going up only 3 levels (../../..) resolves to openseek/competition/LongContext-ICL-Annotation, which makes FLAGOS_ROOT resolve to openseek/competition/LongContext-ICL-Annotation/.. (i.e., openseek/competition). This causes FLAGSCALE_DIR to default to openseek/competition/FlagScale and MODEL_DIR to openseek/competition/models/Qwen3-4B, both of which are incorrect and will cause the script to fail. To reach the actual OpenSeek repository root, you need to go up 6 levels, and 7 levels for FLAGOS_ROOT.

Suggested change

	OPENSEEK_ROOT="$(cd "$COMP_DIR/../../.." && pwd)"
	FLAGOS_ROOT="$(cd "$OPENSEEK_ROOT/.." && pwd)"
	OPENSEEK_ROOT="$(cd "$COMP_DIR/../../../../../.." && pwd)"
	FLAGOS_ROOT="$(cd "$OPENSEEK_ROOT/.." && pwd)"

[gemini-code-assist]

The default versions for PyTorch (2.9.0), torchvision (0.24.0), and torchaudio (2.9.0) do not exist on the PyTorch index and will cause installation failures. Please use valid existing stable versions (e.g., PyTorch 2.5.1, torchvision 0.20.1, and torchaudio 2.5.1).

Suggested change

	TORCH_VERSION="${TORCH_VERSION:-2.9.0}"
	TORCHVISION_VERSION="${TORCHVISION_VERSION:-0.24.0}"
	TORCHAUDIO_VERSION="${TORCHAUDIO_VERSION:-2.9.0}"
	TORCH_VERSION="${TORCH_VERSION:-2.5.1}"
	TORCHVISION_VERSION="${TORCHVISION_VERSION:-0.20.1}"
	TORCHAUDIO_VERSION="${TORCHAUDIO_VERSION:-2.5.1}"

[gemini-code-assist]

The pip install command uses non-existent versions of transformers and openai, which will fail during runtime. Replacing them with valid version constraints prevents execution failures.

Suggested change

	pip install transformers==4.57.6 openai==2.29.0
	pip install "transformers>=4.45.0" "openai>=1.50.0"

[gemini-code-assist]

Several pinned package versions in requirements.txt do not exist on PyPI and will cause installation failures:

• scikit-learn==1.8.0 (latest stable is 1.6.1)
• regex==2025.10.22 (future date-based version)
• openai==2.29.0 (latest stable is 1.x range)
• transformers==4.57.6 (latest stable is 4.48.x range)
• pypdf==5.4.0 (latest stable is 5.2.0 range)

Using looser version constraints or valid existing versions will prevent installation errors.

scikit-learn>=1.5.0
pandas==2.2.3
matplotlib==3.9.4
tqdm==4.67.1
requests==2.32.3
aiohttp==3.11.11
PyYAML==6.0.2
regex>=2024.0.0
typer>=0.9.0
openai>=1.50.0
transformers>=4.45.0
pypdf>=5.0.0

[gemini-code-assist]

Similar to requirements.txt, the pinned versions for scikit-learn, regex, openai, and transformers in environment_h20.yml do not exist on PyPI and will cause conda env create to fail. Using looser constraints or valid existing versions will ensure a successful environment build.

      - scikit-learn>=1.5.0
      - pandas==2.2.3
      - matplotlib==3.9.4
      - tqdm==4.67.1
      - requests==2.32.3
      - aiohttp==3.11.11
      - pyyaml==6.0.2
      - regex>=2024.0.0
      - typer>=0.9.0
      - openai>=1.50.0
      - transformers>=4.45.0