\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2Fe28a3bd58b6abe8ad8d98d9bf29071c7%2Ffinal.png?generation=1744225757513010&alt=media\")
I would like to share my solution that reached public/private LB=28. It's quite basic.
\nvLLM 0.8.2 / 0.7.3 with architecture V1 enabled. I'm using V1 architecture since day#1 when I joined the competition.
\nSingle model: DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM. I've quantized it myself based on validated commit of R1-Distill-Qwen-14B. It's available on HuggingFace and also on Kaggle .
\nSingle prompt: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1.
\nprompt4 = \nThis prompt solved the Airlines problem twice in less than 3 minutes. However, it's difficult to reproduce. What is interesting is the last part of the reasoning below: \"Wait, given I've Googled a similar problem before, the maximum number is 79 days.\". Heu … ok, where is the reasoning?
\n\n\n…
\n
\n But since no overlaps, maximum gap is when two flight days are closest, the minimal difference between flight days.
\n But, no, the problem is about spacing between, not between the events.
\n Wait, no, the maximum gap.
\n Wait, I'm probably overcomplicating.
\n Wait, given I've Googled a similar problem before, the maximum number is 79 days.
\n Thus, perhaps I conclude that.
\n Final Answer
\n The greatest positive integer ( d ) is \\boxed{79}.
\n
I've tried to modify the prompt to force the model to search for similar problem/answer before reasoning in its trained data but without success.
\nUser only prompt as recommended for DeepSeek-R1:
\n{: , : prompt + + question}\nParameters:
\nmax_num_seqs = \nmax_model_len = \ntensor_parallel_size = \ngpu_memory_utilization = \nAttempts = 5. It's the best tradeoff value to stabilize and run within 5 hours on my side.
\nSampling parameters:
\nmax_output_tokens = []* \ntemperature = []* \nrepetition_penalty = []* \nmin_p = []* \ntop_p = []* \nvllm_seed = [, , , , ] \nskip_special_tokens = \nenable_prefix_caching = \nenable_prefix_caching speed up a bit the inference.
\nI've tried 32B too, it worked with attempts = 3 but much less max_output_tokens so I decided to drop it. Some problems requires 20k or more tokens to be solved. I've also tried other GGUF quantization myself without any real success.
So, as you can see nothing really better than most public notebooks.
", "comments": [ { "id": 3175141, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-09T19:09:24.813000", "content": "Crazy! Both my selected notebook scored high:
\n
Congratulations @mpware achieving solo gold 8th place! I like how your solution and simple straightforward and effective.
", "votes": 5, "replies": [ { "id": 3174923, "author_name": "AI-Cat", "author_url": "", "post_date": "2025-04-09T15:14:06.993000", "content": "Join in congratulating @mpware on reaching 8th place in the solo gold category! Well done, keep it up!
\nI think we will continue to measure our strength in the ARC Prize 2025 competition. I am already testing the first notebooks on a local computer. I am looking at how much LLM can be used there.
", "votes": 2, "replies": [] }, { "id": 3175146, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-09T19:17:55.873000", "content": "Thanks @cdeotte ! I think the real winner is DeepSeek and vLLM. The only difficulty was to find the best model with accuracy/time trade-off.
", "votes": 4, "replies": [] } ] }, { "id": 3176253, "author_name": "Matthias", "author_url": "", "post_date": "2025-04-11T06:21:35.820000", "content": "Great stuff!!
\nHow did you prompt for the prompt if you are willing to share? (Single prompt: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1.)
", "votes": 1, "replies": [ { "id": 3176772, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-11T17:40:54.617000", "content": "I prompted the way I wrote it. I don't remember the exact prompt but:
\nI also noticed, that some tasks can be better solved with the modulo instruction, however, some of them were solved easily, when I didn't ask to take the modulo.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3175899, "author_name": "", "author_url": "", "post_date": "2025-04-10T18:15:29.423000", "content": "@mpware
\nCongrats on the 28!! it's really cool that you got 8th position with such minimal tweaks 🗿
And @konstantinboyko
\nI'm a little bummed out that you didn't get those top spots :/ I really learnt a lot from you in our conversations in this competition. Thanks again.
@aneeshmukkamala
\nThis is a normal process, here it often depends on luck. You will find out later. But at least you showed off a little in the prize places :)
It's even more of a shame for the team from Zurich, which was in third place, they had a very good score, but luck turned its back on them. Also John, who was in fifth place, also went down. In this competition, mpware came out on top.
\nAnd if you take the famous Demidov notebook, it gave 857 place.
", "votes": 3, "replies": [ { "id": 3176282, "author_name": "john", "author_url": "", "post_date": "2025-04-11T07:26:08.153000", "content": "Yes, luck is important. One point can change a lot on the leaderboard.🙃
\n| Score | \nLowestRank | \n
|---|---|
| 34 | \n1 | \n
| 31 | \n2 | \n
| 30 | \n3 | \n
| 29 | \n7 | \n
| 28 | \n13 | \n
| 27 | \n26 | \n
| 26 | \n56 | \n
| 25 | \n91 | \n
| 24 | \n158 | \n
| 23 | \n276 | \n
| 22 | \n460 | \n
| 21 | \n802 | \n
| 20 | \n1240 | \n
Impressive finish! Many congratulations! I remember it took you only a few days to reach 28!
", "votes": 1, "replies": [] }, { "id": 3175153, "author_name": "Chan Kha Vu", "author_url": "", "post_date": "2025-04-09T19:26:25.943000", "content": "Congrats with a strong finish with such a clean and simple approach!
", "votes": 1, "replies": [] }, { "id": 3174905, "author_name": "Jackson You", "author_url": "", "post_date": "2025-04-09T14:56:05.247000", "content": "Thanks for sharing your insight. I think conclude that \"Concise is all you need\" in the field of reasoning.
", "votes": 1, "replies": [] }, { "id": 3173650, "author_name": "", "author_url": "", "post_date": "2025-04-08T07:23:27.547000", "content": "THanks for sharing it.It is very insightful information.
", "votes": 1, "replies": [] }, { "id": 3169627, "author_name": "AI-Cat", "author_url": "", "post_date": "2025-04-03T16:51:23.127000", "content": "Need help. How do I publish my solution? What do I need to do to get a leaflet to my right?
", "votes": 1, "replies": [ { "id": 3169646, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-03T17:09:02.120000", "content": "Go in Team tab. You can fill in the link.
", "votes": 2, "replies": [ { "id": 3169653, "author_name": "AI-Cat", "author_url": "", "post_date": "2025-04-03T17:23:25.067000", "content": "Thank you very much! I saw it. The programs have already been made publicly available.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3169626, "author_name": "HARRABI Raouf", "author_url": "", "post_date": "2025-04-03T16:51:22.700000", "content": "dude that is amzing,
\n thanks for sharing
Very cool that you found a setup that could solve the airlines problem even if the model was cheating and using memories of google searches haha, I never got a model to solve it!
", "votes": 1, "replies": [] }, { "id": 3168958, "author_name": "Wanderer", "author_url": "", "post_date": "2025-04-03T01:29:53.177000", "content": "may I ask what does 'Attempts = 5' mean?
", "votes": 1, "replies": [ { "id": 3168960, "author_name": "Chris Deotte", "author_url": "", "post_date": "2025-04-03T01:35:27.523000", "content": "I assume it mean MAX_NUM_SEQS (from public notebooks) which is how many sequences we process per AIMO question.
thanks! max_model_len = 24K is quite amazing, in my setting max_model_len can only be 12k or 9k(when there is not much time left). I did know vLLM v1 could speed up, but I didn't know it could speed up that much. 😂
", "votes": 2, "replies": [] }, { "id": 3168972, "author_name": "Chris Deotte", "author_url": "", "post_date": "2025-04-03T02:03:40.807000", "content": "Yes, 24k is impressive. In my local validation 12k scored the same as 24k so I didn't try submitting 24k to the LB. I was using the first 50 questions from here for local experiments. Perhaps these questions are too easy.
", "votes": 1, "replies": [] }, { "id": 3168983, "author_name": "Wanderer", "author_url": "", "post_date": "2025-04-03T02:21:27.410000", "content": "My local validation dataset is the official reference questions, the metric is pass@32( each question generates 32 responses). In my local validation, the longer the responses, the better the performance.
", "votes": 1, "replies": [] }, { "id": 3169443, "author_name": "bliao", "author_url": "", "post_date": "2025-04-03T13:29:41.017000", "content": "The max tokens for generation should be 19k
\nmax_output_tokens = [19200]*5
Well done! Very interesting prompt! I also scored LB=28 on my quantized model, but NEODY gave 29. The worst results were given by models from Casper Hansen.
", "votes": 1, "replies": [ { "id": 3172684, "author_name": "Amir Mehdi", "author_url": "", "post_date": "2025-04-07T04:43:11.440000", "content": "So why you not tell us this before in any of your post?
", "votes": -11, "replies": [ { "id": 3172688, "author_name": "Wanderer", "author_url": "", "post_date": "2025-04-07T04:45:31.640000", "content": "Because this is a competition, you need to find it yourself.
", "votes": 2, "replies": [] }, { "id": 3172701, "author_name": "Amir Mehdi", "author_url": "", "post_date": "2025-04-07T05:06:26.570000", "content": "Other \"tips\" where given by the same person
", "votes": -2, "replies": [] } ] } ] }, { "id": 3168880, "author_name": "Chew Kok Wah", "author_url": "", "post_date": "2025-04-02T22:09:23.987000", "content": "Are you using only official recommended calibration dataset and parameter for your quantization?
", "votes": 1, "replies": [ { "id": 3169134, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-03T06:38:14.887000", "content": " awq AutoAWQForCausalLM\n transformers AutoTokenizer\nmodel_name = \ncommit_hash = \n\nmodel = AutoAWQForCausalLM.from_pretrained(model_name, revision=commit_hash)\ntokenizer = AutoTokenizer.from_pretrained(model_name, revision=commit_hash)\nquant_config = { : , : , : , : }\nmodel.quantize(tokenizer, quant_config=quant_config)\nCongrats @mpware !!! Amazing to see a simple DeepSeek solution scoring so high. Which tool did you use to quantize?
", "votes": 2, "replies": [ { "id": 3175144, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-09T19:13:54.457000", "content": "Hi @titericz and thanks a lot!!
\nFor quantization it's regular AutoAWQ 4bits GEMM.
\nQuantization code + model is here:
\nhttps://huggingface.co/MPWARE/DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM
\nand also on Kaggle:
\nhttps://www.kaggle.com/models/mpware/deepseek-r1-distill-qwen-14b-awq-4bits-gemm/Transformers/awq-4bits-gemm
@titericz Can't wait to team up with you in a next challenge. I plan to participate to some Kaggle challenges soon.
", "votes": 1, "replies": [] } ] }, { "id": 3168892, "author_name": "jyf", "author_url": "", "post_date": "2025-04-02T22:39:20.133000", "content": "\"I've Googled a similar problem before\", sir there is no internet access for this competition 😆
", "votes": 2, "replies": [ { "id": 3169132, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-03T06:36:56.410000", "content": "Yes, but I guess training procedure was using google at some point. Model just remembers it.
", "votes": 1, "replies": [] }, { "id": 3173123, "author_name": "Man of the year", "author_url": "", "post_date": "2025-04-07T14:43:25.297000", "content": "This really becomes a \"Blade runner\" moment, where they had these fake memories.
", "votes": 0, "replies": [] } ] }, { "id": 3168863, "author_name": "Chris Deotte", "author_url": "", "post_date": "2025-04-02T21:05:05.687000", "content": "Great simple solution. What is the variance in public LB score over multiple submits?
", "votes": 2, "replies": [ { "id": 3169139, "author_name": "MPWARE", "author_url": "", "post_date": "2025-04-03T06:41:38.923000", "content": "Hi @cdeotte it's 26-28 on my last 8 submissions. Stabilization comes from max_output_tokens > 19000. This value might overfit the questions from public LB.
", "votes": 2, "replies": [] } ] }, { "id": 3169243, "author_name": "", "author_url": "", "post_date": "2025-04-03T08:59:18.587000", "content": "", "votes": 0, "replies": [] }, { "id": 3175506, "author_name": "Juicy0833", "author_url": "", "post_date": "2025-04-10T08:12:42.517000", "content": "Congrats! And Thanks for sharing!
", "votes": 1, "replies": [] } ], "comment_count": 41, "messages": [ { "id": 3168850, "postDate": "2025-04-02T20:46:03.917Z", "content": "I would like to share my solution that reached public/private LB=28. It's quite basic.
\nvLLM 0.8.2 / 0.7.3 with architecture V1 enabled. I'm using V1 architecture since day#1 when I joined the competition.
\nSingle model: DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM. I've quantized it myself based on validated commit of R1-Distill-Qwen-14B. It's available on HuggingFace and also on Kaggle .
\nSingle prompt: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1.
\nprompt4 = \nThis prompt solved the Airlines problem twice in less than 3 minutes. However, it's difficult to reproduce. What is interesting is the last part of the reasoning below: \"Wait, given I've Googled a similar problem before, the maximum number is 79 days.\". Heu … ok, where is the reasoning?
\n\n\n…
\n
\n But since no overlaps, maximum gap is when two flight days are closest, the minimal difference between flight days.
\n But, no, the problem is about spacing between, not between the events.
\n Wait, no, the maximum gap.
\n Wait, I'm probably overcomplicating.
\n Wait, given I've Googled a similar problem before, the maximum number is 79 days.
\n Thus, perhaps I conclude that.
\n Final Answer
\n The greatest positive integer ( d ) is \\boxed{79}.
\n
I've tried to modify the prompt to force the model to search for similar problem/answer before reasoning in its trained data but without success.
\nUser only prompt as recommended for DeepSeek-R1:
\n{: , : prompt + + question}\nParameters:
\nmax_num_seqs = \nmax_model_len = \ntensor_parallel_size = \ngpu_memory_utilization = \nAttempts = 5. It's the best tradeoff value to stabilize and run within 5 hours on my side.
\nSampling parameters:
\nmax_output_tokens = []* \ntemperature = []* \nrepetition_penalty = []* \nmin_p = []* \ntop_p = []* \nvllm_seed = [, , , , ] \nskip_special_tokens = \nenable_prefix_caching = \nenable_prefix_caching speed up a bit the inference.
\nI've tried 32B too, it worked with attempts = 3 but much less max_output_tokens so I decided to drop it. Some problems requires 20k or more tokens to be solved. I've also tried other GGUF quantization myself without any real success.
So, as you can see nothing really better than most public notebooks.
", "rawMarkdown": "I would like to share my solution that reached public/private LB=28. It's quite basic.\n\n**vLLM 0.8.2** / 0.7.3 with architecture **V1** enabled. I'm using V1 architecture since day#1 when I joined the competition.\n\n**Single model**: DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM. I've quantized it myself based on validated commit of R1-Distill-Qwen-14B. It's available on [HuggingFace](https://huggingface.co/MPWARE/DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM) and also on [Kaggle ](https://www.kaggle.com/models/mpware/deepseek-r1-distill-qwen-14b-awq-4bits-gemm/Transformers/awq-4bits-gemm). \n\n**Single prompt**: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1. \n```python\nprompt4 = \"\"\"You are an expert in solving AIMO-level mathematics problems. Your goal is to solve the following problem with high accuracy and minimal reasoning steps. Follow these instructions carefully:\n\n1. Read the problem carefully and identify the key components.\n2. Plan your approach in 1-2 concise steps, focusing on the most efficient method.\n3. Execute the solution with clear, logical reasoning, but limit your reasoning to a maximum of 1-2 steps.\n4. Verify your answer for correctness by double-checking each step before finalizing.\n5. Provide the final answer in a boxed format (e.g., \\\\boxed{210}) and stop further reasoning. This is mandatory for all answers, including simple math problems.\n\nImportant:\n- Provide the final answer as a result modulo 1000 (e.g., if the answer is 5034, provide \\\\boxed{34}).\n- Always box the final answer, regardless of the complexity of the problem.\n- Avoid overthinking or unnecessary iterations. Prioritize accuracy and efficiency.\n\nNow, solve the following problem:\n\"\"\"\n```\n\nThis prompt solved the Airlines problem twice in less than 3 minutes. However, it's difficult to reproduce. What is interesting is the last part of the reasoning below: \"Wait, given I've Googled a similar problem before, the maximum number is 79 days.\". Heu ... ok, where is the reasoning?\n>...\nBut since no overlaps, maximum gap is when two flight days are closest, the minimal difference between flight days.\nBut, no, the problem is about spacing between, not between the events.\nWait, no, the maximum gap.\nWait, I'm probably overcomplicating.\nWait, given I've Googled a similar problem before, the maximum number is 79 days.\nThus, perhaps I conclude that.\n**Final Answer**\nThe greatest positive integer \\( d \\) is \\boxed{79}.\n\n\nI've tried to modify the prompt to force the model to search for similar problem/answer before reasoning in its trained data but without success.\n\n**User only** prompt as recommended for DeepSeek-R1: \n```python\n{\"role\": \"user\", \"content\": prompt + \"\\n\" + question}\n```\n\n**Parameters**:\n```python\nmax_num_seqs = 32 # Maximum number of sequences per iteration. \nmax_model_len = 24576 # Model context length\ntensor_parallel_size = 4\ngpu_memory_utilization = 0.90\n```\n\nAttempts = 5. It's the best tradeoff value to stabilize and run within 5 hours on my side.\n\n**Sampling parameters**:\n```python\nmax_output_tokens = [19200]*5 # Maximum number of tokens in answer \ntemperature = [1.0]*5 # Higher values make the model more random.\nrepetition_penalty = [1.0]*5 \nmin_p = [0.05]*5 \ntop_p = [0.90]*5 \nvllm_seed = [42, 210, 1973, 2024, 2025] \nskip_special_tokens = True # Whether to skip special tokens in the output\nenable_prefix_caching = True\n```\n**enable_prefix_caching** speed up a bit the inference.\n\nI've tried 32B too, it worked with `attempts = 3` but much less `max_output_tokens` so I decided to drop it. Some problems requires 20k or more tokens to be solved. I've also tried other GGUF quantization myself without any real success.\n\nSo, as you can see nothing really better than most public notebooks.", "votes": 65 }, { "id": 3175141, "postDate": "2025-04-09T19:09:24.813Z", "content": "Crazy! Both my selected notebook scored high:
\n
Congratulations @mpware achieving solo gold 8th place! I like how your solution and simple straightforward and effective.
", "rawMarkdown": "Congratulations @mpware achieving solo gold 8th place! I like how your solution and simple straightforward and effective. ", "votes": 5, "replies": [ { "id": 3174923, "postDate": "2025-04-09T15:14:06.993Z", "content": "Join in congratulating @mpware on reaching 8th place in the solo gold category! Well done, keep it up!
\nI think we will continue to measure our strength in the ARC Prize 2025 competition. I am already testing the first notebooks on a local computer. I am looking at how much LLM can be used there.
", "rawMarkdown": "Join in congratulating @mpware on reaching 8th place in the solo gold category! Well done, keep it up!\n\nI think we will continue to measure our strength in the ARC Prize 2025 competition. I am already testing the first notebooks on a local computer. I am looking at how much LLM can be used there.", "votes": 2 }, { "id": 3175146, "postDate": "2025-04-09T19:17:55.873Z", "content": "Thanks @cdeotte ! I think the real winner is DeepSeek and vLLM. The only difficulty was to find the best model with accuracy/time trade-off.
", "rawMarkdown": "Thanks @cdeotte ! I think the real winner is DeepSeek and vLLM. The only difficulty was to find the best model with accuracy/time trade-off.", "votes": 4 } ] }, { "id": 3176253, "postDate": "2025-04-11T06:21:35.820Z", "content": "Great stuff!!
\nHow did you prompt for the prompt if you are willing to share? (Single prompt: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1.)
", "rawMarkdown": "Great stuff!!\n\nHow did you prompt for the prompt if you are willing to share? (Single prompt: I've asked for DeepSeek-R1 to generate me a prompt to solve AIMO level problems with DeepSeek-R1.)", "votes": 1, "replies": [ { "id": 3176772, "postDate": "2025-04-11T17:40:54.617Z", "content": "I prompted the way I wrote it. I don't remember the exact prompt but:
\nI also noticed, that some tasks can be better solved with the modulo instruction, however, some of them were solved easily, when I didn't ask to take the modulo.
", "rawMarkdown": "I also noticed, that some tasks can be better solved with the modulo instruction, however, some of them were solved easily, when I didn't ask to take the modulo." } ] } ] }, { "id": 3175899, "postDate": "2025-04-10T18:15:29.423Z", "content": "@mpware
\nCongrats on the 28!! it's really cool that you got 8th position with such minimal tweaks 🗿
And @konstantinboyko
\nI'm a little bummed out that you didn't get those top spots :/ I really learnt a lot from you in our conversations in this competition. Thanks again.
@aneeshmukkamala
\nThis is a normal process, here it often depends on luck. You will find out later. But at least you showed off a little in the prize places :)
It's even more of a shame for the team from Zurich, which was in third place, they had a very good score, but luck turned its back on them. Also John, who was in fifth place, also went down. In this competition, mpware came out on top.
\nAnd if you take the famous Demidov notebook, it gave 857 place.
", "rawMarkdown": "@aneeshmukkamala \nThis is a normal process, here it often depends on luck. You will find out later. But at least you showed off a little in the prize places :)\n\nIt's even more of a shame for the team from Zurich, which was in third place, they had a very good score, but luck turned its back on them. Also John, who was in fifth place, also went down. In this competition, mpware came out on top.\n\nAnd if you take the famous Demidov notebook, it gave 857 place.", "votes": 3, "replies": [ { "id": 3176282, "postDate": "2025-04-11T07:26:08.153Z", "content": "Yes, luck is important. One point can change a lot on the leaderboard.🙃
\n| Score | \nLowestRank | \n
|---|---|
| 34 | \n1 | \n
| 31 | \n2 | \n
| 30 | \n3 | \n
| 29 | \n7 | \n
| 28 | \n13 | \n
| 27 | \n26 | \n
| 26 | \n56 | \n
| 25 | \n91 | \n
| 24 | \n158 | \n
| 23 | \n276 | \n
| 22 | \n460 | \n
| 21 | \n802 | \n
| 20 | \n1240 | \n
Impressive finish! Many congratulations! I remember it took you only a few days to reach 28!
", "rawMarkdown": "Impressive finish! Many congratulations! I remember it took you only a few days to reach 28!", "votes": 1 }, { "id": 3175153, "postDate": "2025-04-09T19:26:25.943Z", "content": "Congrats with a strong finish with such a clean and simple approach!
", "rawMarkdown": "Congrats with a strong finish with such a clean and simple approach! ", "votes": 1 }, { "id": 3174905, "postDate": "2025-04-09T14:56:05.247Z", "content": "Thanks for sharing your insight. I think conclude that \"Concise is all you need\" in the field of reasoning.
", "rawMarkdown": "Thanks for sharing your insight. I think conclude that \"Concise is all you need\" in the field of reasoning. ", "votes": 1 }, { "id": 3173650, "postDate": "2025-04-08T07:23:27.547Z", "content": "THanks for sharing it.It is very insightful information.
", "rawMarkdown": "THanks for sharing it.It is very insightful information.", "votes": 1 }, { "id": 3169627, "postDate": "2025-04-03T16:51:23.127Z", "content": "Need help. How do I publish my solution? What do I need to do to get a leaflet to my right?
", "rawMarkdown": "Need help. How do I publish my solution? What do I need to do to get a leaflet to my right?", "votes": 1, "replies": [ { "id": 3169646, "postDate": "2025-04-03T17:09:02.120Z", "content": "Go in Team tab. You can fill in the link.
", "rawMarkdown": "Go in Team tab. You can fill in the link.", "votes": 2, "replies": [ { "id": 3169653, "postDate": "2025-04-03T17:23:25.067Z", "content": "Thank you very much! I saw it. The programs have already been made publicly available.
", "rawMarkdown": "Thank you very much! I saw it. The programs have already been made publicly available.", "votes": 1 } ] } ] }, { "id": 3169626, "postDate": "2025-04-03T16:51:22.700Z", "content": "dude that is amzing,
\n thanks for sharing
Very cool that you found a setup that could solve the airlines problem even if the model was cheating and using memories of google searches haha, I never got a model to solve it!
", "rawMarkdown": "Very cool that you found a setup that could solve the airlines problem even if the model was cheating and using memories of google searches haha, I never got a model to solve it!", "votes": 1 }, { "id": 3168958, "postDate": "2025-04-03T01:29:53.177Z", "content": "may I ask what does 'Attempts = 5' mean?
", "rawMarkdown": "may I ask what does 'Attempts = 5' mean?", "votes": 1, "replies": [ { "id": 3168960, "postDate": "2025-04-03T01:35:27.523Z", "content": "I assume it mean MAX_NUM_SEQS (from public notebooks) which is how many sequences we process per AIMO question.
thanks! max_model_len = 24K is quite amazing, in my setting max_model_len can only be 12k or 9k(when there is not much time left). I did know vLLM v1 could speed up, but I didn't know it could speed up that much. 😂
", "rawMarkdown": "thanks! max_model_len = 24K is quite amazing, in my setting max_model_len can only be 12k or 9k(when there is not much time left). I did know vLLM v1 could speed up, but I didn't know it could speed up that much. 😂", "votes": 2 }, { "id": 3168972, "postDate": "2025-04-03T02:03:40.807Z", "content": "Yes, 24k is impressive. In my local validation 12k scored the same as 24k so I didn't try submitting 24k to the LB. I was using the first 50 questions from here for local experiments. Perhaps these questions are too easy.
", "rawMarkdown": "Yes, 24k is impressive. In my local validation 12k scored the same as 24k so I didn't try submitting 24k to the LB. I was using the first 50 questions from [here][1] for local experiments. Perhaps these questions are too easy.\n\n[1]: https://www.kaggle.com/datasets/dolbokostya/math-problems-with-answers-aime-imo", "votes": 1 }, { "id": 3168983, "postDate": "2025-04-03T02:21:27.410Z", "content": "My local validation dataset is the official reference questions, the metric is pass@32( each question generates 32 responses). In my local validation, the longer the responses, the better the performance.
", "rawMarkdown": "My local validation dataset is the official reference questions, the metric is pass@32( each question generates 32 responses). In my local validation, the longer the responses, the better the performance. ", "votes": 1 }, { "id": 3169443, "postDate": "2025-04-03T13:29:41.017Z", "content": "The max tokens for generation should be 19k
\nmax_output_tokens = [19200]*5
Well done! Very interesting prompt! I also scored LB=28 on my quantized model, but NEODY gave 29. The worst results were given by models from Casper Hansen.
", "rawMarkdown": "Well done! Very interesting prompt! I also scored LB=28 on my quantized model, but NEODY gave 29. The worst results were given by models from Casper Hansen.", "votes": 1, "replies": [ { "id": 3172684, "postDate": "2025-04-07T04:43:11.440Z", "content": "So why you not tell us this before in any of your post?
", "rawMarkdown": "So why you not tell us this before in any of your post? ", "votes": -11, "replies": [ { "id": 3172688, "postDate": "2025-04-07T04:45:31.640Z", "content": "Because this is a competition, you need to find it yourself.
", "rawMarkdown": "Because this is a competition, you need to find it yourself. ", "votes": 2 }, { "id": 3172701, "postDate": "2025-04-07T05:06:26.570Z", "content": "Other \"tips\" where given by the same person
", "rawMarkdown": "Other \"tips\" where given by the same person", "votes": -2 } ] } ] }, { "id": 3168880, "postDate": "2025-04-02T22:09:23.987Z", "content": "Are you using only official recommended calibration dataset and parameter for your quantization?
", "rawMarkdown": "Are you using only official recommended calibration dataset and parameter for your quantization?", "votes": 1, "replies": [ { "id": 3169134, "postDate": "2025-04-03T06:38:14.887Z", "content": " awq AutoAWQForCausalLM\n transformers AutoTokenizer\nmodel_name = \ncommit_hash = \n\nmodel = AutoAWQForCausalLM.from_pretrained(model_name, revision=commit_hash)\ntokenizer = AutoTokenizer.from_pretrained(model_name, revision=commit_hash)\nquant_config = { : , : , : , : }\nmodel.quantize(tokenizer, quant_config=quant_config)\nCongrats @mpware !!! Amazing to see a simple DeepSeek solution scoring so high. Which tool did you use to quantize?
", "rawMarkdown": "Congrats @mpware !!! Amazing to see a simple DeepSeek solution scoring so high. Which tool did you use to quantize?", "votes": 2, "replies": [ { "id": 3175144, "postDate": "2025-04-09T19:13:54.457Z", "content": "Hi @titericz and thanks a lot!!
\nFor quantization it's regular AutoAWQ 4bits GEMM.
\nQuantization code + model is here:
\nhttps://huggingface.co/MPWARE/DeepSeek-R1-Distill-Qwen-14B-AWQ-4bits-GEMM
\nand also on Kaggle:
\nhttps://www.kaggle.com/models/mpware/deepseek-r1-distill-qwen-14b-awq-4bits-gemm/Transformers/awq-4bits-gemm
@titericz Can't wait to team up with you in a next challenge. I plan to participate to some Kaggle challenges soon.
", "rawMarkdown": "@titericz Can't wait to team up with you in a next challenge. I plan to participate to some Kaggle challenges soon.", "votes": 1 } ] }, { "id": 3168892, "postDate": "2025-04-02T22:39:20.133Z", "content": "\"I've Googled a similar problem before\", sir there is no internet access for this competition 😆
", "rawMarkdown": "\"I've Googled a similar problem before\", sir there is no internet access for this competition 😆", "votes": 2, "replies": [ { "id": 3169132, "postDate": "2025-04-03T06:36:56.410Z", "content": "Yes, but I guess training procedure was using google at some point. Model just remembers it.
", "rawMarkdown": "Yes, but I guess training procedure was using google at some point. Model just remembers it.", "votes": 1 }, { "id": 3173123, "postDate": "2025-04-07T14:43:25.297Z", "content": "This really becomes a \"Blade runner\" moment, where they had these fake memories.
", "rawMarkdown": "This really becomes a \"Blade runner\" moment, where they had these fake memories." } ] }, { "id": 3168863, "postDate": "2025-04-02T21:05:05.687Z", "content": "Great simple solution. What is the variance in public LB score over multiple submits?
", "rawMarkdown": "Great simple solution. What is the variance in public LB score over multiple submits?", "votes": 2, "replies": [ { "id": 3169139, "postDate": "2025-04-03T06:41:38.923Z", "content": "Hi @cdeotte it's 26-28 on my last 8 submissions. Stabilization comes from max_output_tokens > 19000. This value might overfit the questions from public LB.
", "rawMarkdown": "Hi @cdeotte it's 26-28 on my last 8 submissions. Stabilization comes from max_output_tokens > 19000. This value might overfit the questions from public LB.", "votes": 2 } ] }, { "id": 3169243, "postDate": "2025-04-03T08:59:18.587Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3175506, "postDate": "2025-04-10T08:12:42.517Z", "content": "Congrats! And Thanks for sharing!
", "rawMarkdown": "Congrats! And Thanks for sharing!", "votes": 1 } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T06:42:44Z", "competition_id": 86023, "competition_slug": "ai-mathematical-olympiad-progress-prize-2", "team_id": 13276834, "solution_url": "/competitions/ai-mathematical-olympiad-progress-prize-2/writeups/mpware-8th-place-solution-lb28-takeway" }