{ "id": 573071, "title": "[Private27,Public28] Solution for 17th: To get code & fix it with DeepSeek", "content": "

LLM used:
\n deepseek-r1-distill-qwen-14b-awq-casperhansen

\n

I thought some of the questions can be solved much more efficiently by running Python code— so I explored generating code with DeepSeek R1. However, simply adding \"Make python code.\" to the prompt did not work well.
\nHere is what I found that worked well with DeepSeek and QWQ,

\n

To get code and fix it with LLM:

\n

Step 1: Initial Prompting

\n

Provide the question text as a prompt to the model.
\nSet output length between 2000 to 4000 tokens.
\nWith this, get output in 1-2 min.
\nOutput text are always truncated but no problem for further use for code generation.

\n

Step 2: Prompt for Code Generation

\n

Construct a new retry_prompt using the model's earlier output:
\nStructure:

\n
    content = <Instruction text  make code  previous thought>\n     (e.g.)\n    content += question\n    content += output\n\n    prompt = [{: , : content}]\n    prompt = tokenizer.apply_chat_template(\n                        =retry_prompt,\n                        =,\n                        =)\n
\n

Then, add either of the following:

\n

For better code (more verbose, with reasoning):

\n
    prompt += '</think>'\n    prompt += \"Here's a structured approach to solve the problem:### Approach\"\n
\n

For fast (but sometimes lower-quality) code:

\n
     += \n    prompt += \n
\n

With this, get code within another ~1 min.

\n

Step 3: Error Fixing Prompt

\n

If generated code fails when run via subprocess, build a new error-fixing prompt:

\n
    content = <Some instructions to fix the code>\n     (e.g.>\n    content +=  question\n    content += <Code with error>\n    content += <Error message>\n
\n

Retry generation with this new prompt.
\nSometime works, but sometimes encounters another error and repeat this fixing process.

\n

Notebook demonstrating code generation and error fixing using the reference set:
\nhttps://www.kaggle.com/code/ippeiogawa/generate-code-from-deepseekr1

\n

Summary of strategy

\n\n

Thank you for reading!

", "comments": [ { "id": 3177995, "author_name": "Md Adilur Rahim", "author_url": "", "post_date": "2025-04-13T15:59:47.333000", "content": "

I did the same thing. I found that the Deepseek R1 Distil 32 B model is great for Python coding. My approach failed the public and private sets, and maybe needed more evaluations and the early stopping tactics.

\n

I was confident that the code-generating solution would work. I am so glad to see your approach worked!

", "votes": 2, "replies": [] }, { "id": 3177863, "author_name": "ducnh279", "author_url": "", "post_date": "2025-04-13T11:11:56.990000", "content": "

Congrats, I really like your approach!

", "votes": 0, "replies": [] } ], "comment_count": 2, "messages": [ { "id": 3177845, "postDate": "2025-04-13T10:42:31.920Z", "content": "

LLM used:
\n deepseek-r1-distill-qwen-14b-awq-casperhansen

\n

I thought some of the questions can be solved much more efficiently by running Python code— so I explored generating code with DeepSeek R1. However, simply adding \"Make python code.\" to the prompt did not work well.
\nHere is what I found that worked well with DeepSeek and QWQ,

\n

To get code and fix it with LLM:

\n

Step 1: Initial Prompting

\n

Provide the question text as a prompt to the model.
\nSet output length between 2000 to 4000 tokens.
\nWith this, get output in 1-2 min.
\nOutput text are always truncated but no problem for further use for code generation.

\n

Step 2: Prompt for Code Generation

\n

Construct a new retry_prompt using the model's earlier output:
\nStructure:

\n
    content = <Instruction text  make code  previous thought>\n     (e.g.)\n    content += question\n    content += output\n\n    prompt = [{: , : content}]\n    prompt = tokenizer.apply_chat_template(\n                        =retry_prompt,\n                        =,\n                        =)\n
\n

Then, add either of the following:

\n

For better code (more verbose, with reasoning):

\n
    prompt += '</think>'\n    prompt += \"Here's a structured approach to solve the problem:### Approach\"\n
\n

For fast (but sometimes lower-quality) code:

\n
     += \n    prompt += \n
\n

With this, get code within another ~1 min.

\n

Step 3: Error Fixing Prompt

\n

If generated code fails when run via subprocess, build a new error-fixing prompt:

\n
    content = <Some instructions to fix the code>\n     (e.g.>\n    content +=  question\n    content += <Code with error>\n    content += <Error message>\n
\n

Retry generation with this new prompt.
\nSometime works, but sometimes encounters another error and repeat this fixing process.

\n

Notebook demonstrating code generation and error fixing using the reference set:
\nhttps://www.kaggle.com/code/ippeiogawa/generate-code-from-deepseekr1

\n

Summary of strategy

\n\n

Thank you for reading!

", "rawMarkdown": "LLM used:\n [deepseek-r1-distill-qwen-14b-awq-casperhansen](https://www.kaggle.com/models/huikang/deepseek-r1/Transformers/deepseek-r1-distill-qwen-14b-awq-casperhansen/1)\n\nI thought some of the questions can be solved much more efficiently by running Python code— so I explored generating code with DeepSeek R1. However, simply adding \"Make python code.\" to the prompt did not work well.\nHere is what I found that worked well with DeepSeek and QWQ,\n#To get code and fix it with LLM:\nStep 1: Initial Prompting\n\nProvide the question text as a prompt to the model.\nSet output length between 2000 to 4000 tokens. \nWith this, get output in 1-2 min.\nOutput text are always truncated but no problem for further use for code generation.\n\nStep 2: Prompt for Code Generation\n\nConstruct a new retry_prompt using the model's earlier output:\nStructure:\n\n content = \n (e.g.\"Please make a short summary of your approach, including python code.\")\n content += question\n content += output\n\n prompt = [{\"role\": \"user\", \"content\": content}]\n prompt = tokenizer.apply_chat_template(\n conversation=retry_prompt,\n tokenize=False,\n add_generation_prompt=True)\n\nThen, add either of the following:\n\nFor better code (more verbose, with reasoning):\n\n prompt += '\\n\\n'\n prompt += \"Here's a structured approach to solve the problem:\\n### Approach\\n\"\n\nFor fast (but sometimes lower-quality) code:\n\n prompt += '\\n\\n'\n prompt += '```python\\n'\n\nWith this, get code within another ~1 min.\n\nStep 3: Error Fixing Prompt\n\nIf generated code fails when run via subprocess, build a new error-fixing prompt:\n\n content = \n (e.g.\"Please correct python code below and put in ```python\\n{{corrected code here}}```.\">\n content += question\n content += \n content += \n\nRetry generation with this new prompt.\nSometime works, but sometimes encounters another error and repeat this fixing process.\n\nNotebook demonstrating code generation and error fixing using the reference set:\n[https://www.kaggle.com/code/ippeiogawa/generate-code-from-deepseekr1](https://www.kaggle.com/code/ippeiogawa/generate-code-from-deepseekr1)\n\n\n\n#Summary of strategy\n- Samples:\nTotal of 16 samples —\n→ 4 used to get code (as mentioned above)\n→ 12 used to get boxed answers\n- Chunked Reasoning:\nOutput every 2000 tokens → combine 3–4 outputs → prompt with:\n\"Please think again from combined previous thought.\"\n- vLLM.LLMEngine:\nUsed to stream results as soon as a sample finishes.\nLLMEngine Code Snippet:\n\n from vllm import EngineArgs, LLMEngine, SamplingParams\n\n # Initialize engine\n engine_args = EngineArgs(\n llm_model_path,\n max_model_len=max_len,\n trust_remote_code=True,\n tensor_parallel_size=4,\n gpu_memory_utilization=0.96,\n seed=2025,\n )\n engine = LLMEngine.from_engine_args(engine_args)\n\n # Main loop\n while :\n if prompt:\n engine.add_request(str(request_id), prompt, sampling_params)\n request_id +=1\n #\n request_outputs = engine.step()\n #\n for request_output in request_outputs:\n if request_output.finished:\n output_text = request_output[0].text\n # Extract an answer\n # Process output\n if Code in output_text:\n run via subprocess\n #\n # Build new prompts for generating or fixing code or to think again with combined thought.\n new_prompt = \n engine.add_request(str(request_id), new_prompt, new_sampling_params)\n request_id +=1\n\n- Early stopping: \nIf 3 identical answers are produced, I treat that as the final answer and move on to the next question.\n\nThank you for reading!", "votes": 15 }, { "id": 3177995, "postDate": "2025-04-13T15:59:47.333Z", "content": "

I did the same thing. I found that the Deepseek R1 Distil 32 B model is great for Python coding. My approach failed the public and private sets, and maybe needed more evaluations and the early stopping tactics.

\n

I was confident that the code-generating solution would work. I am so glad to see your approach worked!

", "rawMarkdown": "I did the same thing. I found that the Deepseek R1 Distil 32 B model is great for Python coding. My approach failed the public and private sets, and maybe needed more evaluations and the early stopping tactics. \n\nI was confident that the code-generating solution would work. I am so glad to see your approach worked! ", "votes": 2 }, { "id": 3177863, "postDate": "2025-04-13T11:11:56.990Z", "content": "

Congrats, I really like your approach!

", "rawMarkdown": "Congrats, I really like your approach!" } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T06:42:10Z", "competition_id": 86023, "competition_slug": "ai-mathematical-olympiad-progress-prize-2", "team_id": 12923091, "solution_url": "/competitions/ai-mathematical-olympiad-progress-prize-2/writeups/ippeiogawa-private27-public28-solution-for-17th-to" }