Add files via upload

[quoteven]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces task-specific prompt routing (M03, CoT) and dynamic example selection strategies (M05, M09, M10, M11, M19, M20) to optimize in-context learning (ICL) annotations. It also switches the annotation backend to Huawei Ascend with differentiated temperature and token configurations. However, several critical issues were identified: caching examples_str defeats dynamic retrieval and breaks the mixed context length strategy; a typo in the regex pattern (<think&rt;) prevents thinking process removal; hardcoded absolute paths limit portability; unsafe indexing of example['output'][0] can truncate string outputs; repeatedly initializing the tokenizer creates a performance bottleneck; and the calculate_similarity function is duplicated across multiple methods.

[gemini-code-assist]

Caching examples_str with if examples_str is None: completely defeats the purpose of dynamic retrieval (e.g., select_examples_M11 which selects examples based on the current text2annotate). It also breaks the mixed context length strategy: since sample_idx = 0 uses a 30k context, the selected 30k token examples will be reused for all subsequent samples, bypassing the 8k context efficiency optimization for sample_idx >= 50 and potentially causing out-of-memory or context limit issues.

        # M11优化：使用Task 7 Jeopardy线索拆解策略
        examples_str = select_examples_M11(icl_examples, task_description, text2annotate, task_id, sample_idx)

[gemini-code-assist]

The regex pattern contains a typo <think&rt; and </think&rt;. Standard reasoning models output <think> and </think>. This typo will prevent the thinking process from being matched and removed, leading to incorrect code extraction or parsing failures.

Suggested change

	# Step 1: 过滤<think&rt;标签内容（去思维链）
	cleaned_result = re.sub(r'<think&rt;.*?</think&rt;', '', whole_result, flags=re.DOTALL)
	# Step 1: 过滤<think>标签内容（去思维链）
	cleaned_result = re.sub(r'<think>.*?</think>', '', whole_result, flags=re.DOTALL)

[gemini-code-assist]

Hardcoding absolute paths like /root/OpenSeek/... makes the codebase non-portable and will cause FileNotFoundError on other environments. It is highly recommended to resolve these paths dynamically relative to the script's location.

_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
_DATA_DIR = os.path.join(os.path.dirname(_SCRIPT_DIR), 'data')

TASK_FILES = {
    1: os.path.join(_DATA_DIR, 'openseek-1_closest_integers.json'),
    2: os.path.join(_DATA_DIR, 'openseek-2_count_nouns_verbs.json'),
    3: os.path.join(_DATA_DIR, 'openseek-3_collatz_conjecture.json'),
    4: os.path.join(_DATA_DIR, 'openseek-4_conala_concat_strings.json'),
    5: os.path.join(_DATA_DIR, 'openseek-5_semeval_2018_task1_tweet_sadness_detection.json'),
    6: os.path.join(_DATA_DIR, 'openseek-6_mnli_same_genre_classification.json'),
    7: os.path.join(_DATA_DIR, 'openseek-7_jeopardy_answer_generation_all.json'),
    8: os.path.join(_DATA_DIR, 'openseek-8_kernel_generation.json'),
}

[gemini-code-assist]

Accessing example['output'][0] assumes that example['output'] is always a list. If example['output'] is a string (e.g., "Good Review"), this will silently return only the first character (e.g., "G"), resulting in truncated labels in the prompt. This issue also exists in select_examples_M19, select_examples_M09, select_examples_M10, and select_examples_M11.

Suggested change

	input_text = example['input']
	output_text = example['output'][0]
	input_text = example['input']
	output_text = example['output'][0] if isinstance(example['output'], list) else example['output']

[gemini-code-assist]

Initializing the tokenizer from disk inside the example selection function will cause a massive performance bottleneck if called per-sample during dynamic retrieval. It is highly recommended to use a global tokenizer singleton (similar to get_tokenizer_m20) or pass the pre-loaded tokenizer as an argument.

Suggested change

	# 初始化tokenizer
	tokenizer = AutoTokenizer.from_pretrained("/root/Qwen3-4B", trust_remote_code=True)
	# 使用全局tokenizer单例避免重复加载
	tokenizer = get_tokenizer_m20()

[gemini-code-assist]

The calculate_similarity function is defined identically inside select_examples_M05, select_examples_M20 (as calculate_similarity_m20), and select_examples_M19 (as calculate_similarity_m19). Defining it once at the module level will improve maintainability and reduce code duplication.