feat: add new long-context ICL annotation method

[xiamaozi11]

No description provided.

[gemini-code-assist]

Code Review

This pull request implements a comprehensive automatic data annotation solution for the OpenSeek competition (Tasks 1-8), featuring hybrid ICL retrieval, spaCy-based POS tagging, and ReAct-style verification loops. The code review feedback highlights critical security concerns regarding hardcoded DeepSeek API keys in build_cot_training_set.py and deepseek_client.py. Additionally, there are opportunities to improve portability by removing hardcoded CUDA_VISIBLE_DEVICES settings, clean up leftover debugging print statements in the Task 5 inference scripts, and enhance the robustness of the output comparison script by supporting alternative field names.

[gemini-code-assist]

Hardcoding API keys is a severe security risk. It is highly recommended to remove the hardcoded key and raise an error if the DEEPSEEK_API_KEY environment variable is not set.

Suggested change

	if not args.dry_run and not api_key:
	api_key = 'sk-340c73f8209e4ba3ae44c7e6cf570506'
	if not args.dry_run and not api_key:
	raise ValueError("DEEPSEEK_API_KEY environment variable is not set.")

[gemini-code-assist]

Hardcoding API keys is a severe security risk. Please retrieve the API key from the environment variable DEEPSEEK_API_KEY instead.

Suggested change

	DEEPSEEK_API_KEY = "sk-340c73f8209e4ba3ae44c7e6cf570506"
	DEEPSEEK_API_KEY = os.environ.get("DEEPSEEK_API_KEY", "")

[gemini-code-assist]

This print statement appears to be a leftover debugging aid. It should be removed to keep the output clean.

[gemini-code-assist]

This print statement appears to be a leftover debugging aid. It should be removed to keep the output clean.

[gemini-code-assist]

The current implementation only supports test_sample_id and prediction fields. Supporting both test_sample_id/example_id and prediction/model_output makes this comparison script much more robust and compatible with both examples and test samples.

Suggested change

	sample_id = obj.get("test_sample_id")
	prediction = obj.get("prediction")
	if sample_id is None or prediction is None:
	raise ValueError(
	f"{path} 第 {line_no} 行缺少字段，必须包含 test_sample_id 和 prediction"
	)
	sample_id = obj.get("test_sample_id") or obj.get("example_id")
	prediction = obj.get("prediction") or obj.get("model_output")
	if sample_id is None or prediction is None:
	raise ValueError(
	f"{path} line {line_no} is missing required fields (must contain test_sample_id/example_id and prediction/model_output)"
	)

[gemini-code-assist]

Hardcoding CUDA_VISIBLE_DEVICES to a specific index like "2" limits portability and can cause execution failures on systems with fewer GPUs. It is better to let the user configure this via their environment.