1️⃣ CoT Infrastructure & Sentence Taxonomy

# Configuration
MODEL_NAME_1B = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
MODEL_NAME_8B = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
MODEL_NAME_14B = "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B"
DATASET_NAME = "uzaymacar/math-rollouts"
BLACKMAIL_DATASET_NAME = "uzaymacar/blackmail-rollouts"
EMBEDDING_MODEL = "all-MiniLM-L6-v2"
SIMILARITY_THRESHOLD = 0.8

embedding_model = SentenceTransformer(EMBEDDING_MODEL)
print(embedding_model)

prompts = [
    "Wait, I think I made an error in my reasoning and need to backtrack",
    "Hold on, I believe I made a mistake in my logic and should reconsider",
    "After careful analysis, I've determined the correct answer is 42",
    "Time is an illusion. Lunchtime doubly so.",
]
labels = [x[:35] + "..." for x in prompts]

embedding = embedding_model.encode(prompts)
cosine_sims = embedding @ embedding.T

fig, ax = plt.subplots(figsize=(8, 6))
im = ax.imshow(cosine_sims, cmap="RdBu", vmin=-1, vmax=1)
ax.set_xticks(range(len(labels)))
ax.set_yticks(range(len(labels)))
ax.set_xticklabels(labels, rotation=45, ha="right", fontsize=9)
ax.set_yticklabels(labels, fontsize=9)
plt.colorbar(im, label="Cosine Similarity")
plt.title("Sentence Embedding Similarity")
plt.tight_layout()
plt.show()

PROBLEM_ID = 4682

path = f"deepseek-r1-distill-llama-8b/temperature_0.6_top_p_0.95/correct_base_solution/problem_{PROBLEM_ID}/base_solution.json"
local_path = hf_hub_download(repo_id=DATASET_NAME, filename=path, repo_type="dataset")

with open(local_path, "r") as f:
    problem_data = json.load(f)

print("Keys in problem data:", list(problem_data.keys()))
print(f"\nProblem prompt:\n{problem_data['prompt']}")

def split_solution_into_chunks(text: str) -> list[str]:
    """Split solution into sentence-level chunks."""
    raise NotImplementedError("Implement sentence splitting")


test_cases = [
    (
        "<think>First, I understand the problem. Next, I'll solve for x. Finally, I verify!</think>",
        ["First, I understand the problem.", "Next, I'll solve for x.", "Finally, I verify!"],
    ),
    (
        "<think>Let me break this down: 1. Convert to decimal. 2. Calculate log. 3. Apply formula.</think>",
        ["Let me break this down:", "1. Convert to decimal.", "2. Calculate log.", "3. Apply formula."],
    ),
    (
        "<think>The answer is 42. Done.</think>",
        ["The answer is 42.", "Done."],
    ),
]

for input_text, expected_chunks in test_cases:
    chunks = split_solution_into_chunks(input_text)
    assert chunks == expected_chunks, f"Expected {expected_chunks}, got {chunks}"

print("All tests passed!")

CATEGORIES = {
    "problem_setup": "Problem Setup",
    "plan_generation": "Plan Generation",
    "fact_retrieval": "Fact Retrieval",
    "active_computation": "Active Computation",
    "uncertainty_management": "Uncertainty Management",
    "result_consolidation": "Result Consolidation",
    "self_checking": "Self Checking",
    "final_answer_emission": "Final Answer Emission",
    "unknown": "Unknown",
}

CATEGORY_WORDS = {
    # Fill in keywords for each category
    "final_answer_emission": ["\\boxed", "final answer"],
    "problem_setup": [],  # Add keywords
    "fact_retrieval": [],  # Add keywords
    # ... add more categories
}


def categorize_sentences_heuristic(chunks: list[str]) -> list[str]:
    """Categorize sentences using keyword matching."""
    raise NotImplementedError("Implement heuristic categorization")


example_sentences = [
    "I need to find the area of a circle with radius 5.",
    "The formula for circle area is A = πr².",
    "Substituting r = 5: A = π × 5² = 25π.",
    "Wait, let me look again at that calculation.",
    "So the area is 25π square units.",
    "Therefore, the answer is \\boxed{25π}.",
]
expected_categories = [
    "Problem Setup",
    "Fact Retrieval",
    "Active Computation",
    "Uncertainty Management",
    "Result Consolidation",
    "Final Answer Emission",
]

categories = categorize_sentences_heuristic(example_sentences)
for sentence, cat, expected in zip(example_sentences, categories, expected_categories):
    assert cat == expected, f"Expected {expected!r}, got {cat!r} for: {sentence!r}"

print("All tests for `categorize_sentences_heuristic` passed!")

def generate_response(
    model: str,
    messages: list[dict[str, str]],
    max_tokens: int = 128,
    stop_sequences: list[str] | None = None,
    temperature: float = 1.0,
    max_retries: int = 10,
) -> str:
    """Single API call with retry logic for rate limits."""
    assert openrouter_client, "OpenRouter API key not set (see earlier instructions)."

    stop_sequences = stop_sequences or []

    for attempt in range(max_retries):
        try:
            resp = openrouter_client.chat.completions.create(
                model=model,
                messages=messages,
                max_tokens=max_tokens,
                temperature=temperature,
                stop=stop_sequences if stop_sequences else None,
            )
            return resp.choices[0].message.content or ""
        except Exception as e:
            print(str(e))
            if "rate_limit" in str(e).lower() or "429" in str(e):
                if attempt < max_retries - 1:
                    wait_time = 2**attempt
                    print(f"Rate limit hit, waiting {wait_time}s...")
                    time.sleep(wait_time)
                    continue
            raise e
    return ""


def generate_responses_parallel(
    messages_list: list[list[dict[str, str]]],
    model: str,
    max_tokens: int = 128,
    stop_sequences: list[str] | None = None,
    temperature: float = 1.0,
    max_workers: int = 10,
) -> list[str | Exception]:
    """
    Run multiple API calls in parallel using ThreadPoolExecutor.

    Args:
        messages_list: List of message lists, each to be sent as a separate API call
        model: Model identifier for OpenRouter
        max_tokens: Max tokens per response
        stop_sequences: Stop sequences for generation
        temperature: Sampling temperature
        max_workers: Maximum number of parallel workers

    Returns:
        List of responses (strings) or Exceptions for failed calls, in same order as input
    """
    results: dict[int, str | Exception] = {}
    pbar = tqdm(total=len(messages_list), desc="API calls")

    def call_single(idx: int, messages: list[dict[str, str]]) -> tuple[int, str | Exception]:
        """Helper function to make a single API call."""
        try:
            time.sleep(0.1)  # Rate limiting
            result = generate_response(model, messages, max_tokens, stop_sequences, temperature)
            return idx, result
        except Exception as e:
            return idx, e

    # Execute tasks in parallel
    with ThreadPoolExecutor(max_workers=max_workers) as executor:
        # Submit all tasks with their indices
        futures = [executor.submit(call_single, i, m) for i, m in enumerate(messages_list)]

        # Process completed tasks
        for future in as_completed(futures):
            idx, result = future.result()
            results[idx] = result
            pbar.update(1)

    pbar.close()

    # Return results in original order
    return [results[i] for i in range(len(messages_list))]


# Demo of how this function works:
sys_prompt = {"role": "system", "content": "Reply in rhyming couplets."}
test_messages = [
    [sys_prompt, {"role": "user", "content": "What is 2+2?"}],
    [sys_prompt, {"role": "user", "content": "What is the capital of France?"}],
]
responses = generate_responses_parallel(test_messages, model="openai/gpt-4.1-mini", max_tokens=40)
for i, response in enumerate(responses):
    print(f"Response {i + 1}:\n{response}\n")

DAG_SYSTEM_PROMPT = """
You are an expert in interpreting how language models solve math problems using multi-step reasoning. Your task is to analyze a Chain-of-Thought (CoT) reasoning trace, broken into discrete text chunks, and label each chunk with a tag that describes what this chunk is *doing* functionally in the reasoning process.

### Function Tags:

1. `problem_setup`: Parsing or rephrasing the problem
2. `plan_generation`: Stating or deciding on a plan of action
3. `fact_retrieval`: Recalling facts, formulas, problem details
4. `active_computation`: Performing algebra, calculations, manipulations
5. `result_consolidation`: Aggregating intermediate results, summarizing
6. `uncertainty_management`: Expressing confusion, re-evaluating, backtracking
7. `final_answer_emission`: Explicit statement of the final boxed answer
8. `self_checking`: Verifying previous steps
9. `unknown`: Use only if the chunk does not fit any of the above

### Output Format:

Return a numbered list with one function tag per chunk:

1. problem_setup
2. active_computation
...
"""


def categorize_sentences_autorater(
    problem_text: str, chunks: list[str], model: str = "openai/gpt-4.1-mini"
) -> list[str]:
    """Categorize sentences using an LLM autorater."""
    raise NotImplementedError("Implement autorater categorization")


example_problem = "What is the area of a circle with radius 5?"
autorater_categories = categorize_sentences_autorater(example_problem, example_sentences)

df = pd.DataFrame(
    {
        "Sentence": example_sentences,
        "Heuristic": categories,
        "Autorater": autorater_categories,
    }
)
display(df)