Add OpenSeek submission for LongContext-ICL-Annotation

[hdxdfff]

Summary

• Add the OpenSeek submission for the LongContext-ICL-Annotation challenge.
• Include the final prediction archive, the final technical report, and the final source-code archive used for the competition submission.
• Also include an extracted code/ directory for direct review and reproduction.
• Final confirmed platform score for this submission line: 87.35.

Compliance

• Uses Qwen3-4B as the only large language model.
• Uses FlagScale as the deployment and inference framework.
• Uses only the official competition data.
• Does not rely on fine-tuning, external data, or external models.
• Public materials have been cleaned to remove unnecessary internal diagnostics and account-specific bindings, while keeping the final submitted artifacts unchanged.

Reproduction

• See openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/README.md for the submission overview.
• See openseek/competition/LongContext-ICL-Annotation/submissions/OpenSeek/code/README.md for environment setup, model deployment, inference, packaging, and validation instructions.

[gemini-code-assist]

Code Review

This pull request introduces the complete OpenSeek submission codebase, including configuration files, prompt templates, utility scripts, source code for retrieval, decision-making, and pipeline execution, as well as tests and a technical report. The code review identified several critical issues and improvement opportunities: incorrect dataset paths (RAW_TASK8) in scripts that would cause FileNotFoundErrors, potential runtime NameErrors in fallback code generation, a bug where max_samples being 0 would be treated as falsy, and a logger that creates a directory but fails to write logs to a file. Additionally, feedback was provided to enforce strict error handling when packaging missing prediction files and to clean up unused dead code.

[gemini-code-assist]

The path to RAW_TASK8 is incorrect. According to the directory structure defined in data/README.md and the behavior of import_official_data.sh, the official dataset files are imported into data/raw/openseek/. Therefore, the path should include the openseek subdirectory to avoid a FileNotFoundError when running this script.

Suggested change

	RAW_TASK8 = ROOT / "data/raw/openseek-8_kernel_generation.json"
	RAW_TASK8 = ROOT / "data/raw/openseek/openseek-8_kernel_generation.json"

[gemini-code-assist]

The path to RAW_TASK8 is incorrect. The official dataset files are imported into data/raw/openseek/ by import_official_data.sh. The path should include the openseek subdirectory to avoid a FileNotFoundError when running this script.

Suggested change

	RAW_TASK8 = ROOT / "data/raw/openseek-8_kernel_generation.json"
	RAW_TASK8 = ROOT / "data/raw/openseek/openseek-8_kernel_generation.json"

[gemini-code-assist]

If names is empty (e.g., when args is "*args, **kwargs"), setting first to "input" will cause a NameError during execution because input is not defined in the function signature. Since the parameter is *args, using "args[0]" instead of "input" is a robust fallback that correctly references the first positional argument.

Suggested change

	first = names[0] if names else "input"
	first = names[0] if names else "args[0]"

[gemini-code-assist]

Hardcoding optional parameters like bias, stride, padding, dilation, and groups directly as positional arguments can cause a NameError if they are not defined in the wrapper's signature. Using locals().get(name, default) is a much more robust approach that safely retrieves these parameters if they exist, or falls back to their PyTorch defaults.

Suggested change

	expr = "torch.nn.functional.conv2d(input, weight, bias, stride, padding, dilation, groups)"
	expr = "torch.nn.functional.conv2d(input, weight, locals().get('bias', None), locals().get('stride', 1), locals().get('padding', 0), locals().get('dilation', 1), locals().get('groups', 1))"

[gemini-code-assist]

Packaging the submission when required prediction files are missing will result in an incomplete and invalid submission package. Instead of just logging a warning, the script should raise a FileNotFoundError to fail the packaging process and alert the user immediately.

Suggested change

	if missing:
	logger.warning("Packaging with missing prediction files: %s", ", ".join(missing))
	if missing:
	raise FileNotFoundError(f"Cannot package submission. Missing required prediction files: {', '.join(missing)}")

[gemini-code-assist]

Using if max_samples to check if max_samples is set can lead to incorrect behavior if max_samples_per_task is set to 0 in the configuration (as 0 is falsy in Python, it would fall back to evaluating all samples instead of 0). It is safer and more idiomatic to explicitly check if max_samples is not None.

Suggested change

	sample_records = dataset.test_samples[: int(max_samples)] if max_samples else dataset.test_samples
	sample_records = dataset.test_samples[: int(max_samples)] if max_samples is not None else dataset.test_samples

[gemini-code-assist]

The TokenCounter class and the maybe_build_token_counter function are completely unused in the codebase. Additionally, the chunk_budget_tokens parameter in configs/base.yaml is also unused. Removing dead code and unused configuration parameters improves maintainability and clarity.

[gemini-code-assist]

The build_logger function creates the log_dir directory but only configures a StreamHandler to log to the console. It does not add a FileHandler to actually write the logs to a file in log_dir. Adding a FileHandler ensures that logs are persisted for future diagnostics.

    log_path = Path(log_dir)
    log_path.mkdir(parents=True, exist_ok=True)
    logger = logging.getLogger(name)
    logger.setLevel(logging.INFO)

    if not logger.handlers:
        formatter = logging.Formatter(
            "%(asctime)s | %(levelname)s | %(name)s | %(message)s"
        )
        
        stream_handler = logging.StreamHandler()
        stream_handler.setFormatter(formatter)
        logger.addHandler(stream_handler)
        
        file_handler = logging.FileHandler(log_path / "run.log", encoding="utf-8")
        file_handler.setFormatter(formatter)
        logger.addHandler(file_handler)

[gemini-code-assist]

The functions format_example, build_output_instruction, and build_prompt are unused in the codebase. Removing unused utility functions keeps the codebase clean and maintainable.