| Field | \nValue | \n
|---|---|
| Competition Name | \nAI Mathematical Olympiad - Progress Prize 3 | \n
| Team Name | \nAndreas Bisiadis | \n
| Private Leaderboard Score | \n41.5 | \n
| Private Leaderboard Place | \n546 | \n
| Public Leaderboard Reference | \n41/50 for the selected Mk I Version 16 notebook; 42/50 peak for the public no-suffix notebook lineage | \n
| Team Member Name | \nAndreas Bisiadis | \n
This submission was developed for Kaggle's AI Mathematical Olympiad - Progress Prize 3 competition. The selected solution is an inference-only notebook based on GPT-OSS-120B with tool-integrated Python reasoning. The notebook selected for the final submission was Mk I Version 16:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThis selected notebook scored 41/50 on the public leaderboard and 41.5 on the private leaderboard. The same public, no-suffix notebook lineage reached a peak public score of 42/50 in Version 11.
\nThis was a solo submission by Andreas Bisiadis. The technical work represented in this repository focused on long-context LLM inference systems, vLLM backend tuning, tool-integrated mathematical reasoning, Jupyter kernel sandbox reliability, prompt minimization, and answer aggregation under Kaggle's fixed H100 runtime constraints.
\nPrimary competition work included:
\ngpt-oss-120b-with-tools, through 22 versions. This is the only lineage intended to correspond to the publicly available Kaggle notebook.vllm==0.17.0, while the selected public Mk I Version 16 notebook used vllm==0.11.2.The final selected solution is an inference-time ensemble built around GPT-OSS-120B served locally with vLLM on a single Kaggle H100 GPU. Each math problem is solved by up to 8 parallel attempts using the Harmony prompt format, ReasoningEffort.HIGH, and a Python tool backed by persistent Jupyter kernels. Candidate integer answers are extracted from final answers using a \\boxed{} parser with a narrow textual fallback for \"final answer is\", then aggregated with inverse-entropy weighting computed from top-logprob token entropy. The selected Mk I Version 16 notebook uses vllm==0.11.2, a 65,536-token context, FP8 E4M3 KV cache, async scheduling, disabled log stats, and prefix caching. No model training or fine-tuning is performed; development effort was concentrated on inference robustness, tool reliability, runtime allocation, and answer selection.
The effective inputs and engineered signals are:
\n\\boxed{} format.math, numpy, and sympy.The main transformations are prompt composition, answer normalization, and entropy-based scoring. Boxed answer extraction accepts comma separators, strips them, converts to an integer, and keeps only values in [0, 99999]. Version 16 also includes a narrow case-insensitive fallback for text matching final answer is <integer>.
No external problem data or retrieval corpus is used. The only external artifacts are allowed Kaggle inputs: the competition data, the GPT-OSS-120B model weights, utility wheels, and local tiktoken encoding files needed for offline execution.
\nNo training, fine-tuning, reinforcement learning, or preference optimization is used in the submitted solution. The model is the pretrained GPT-OSS-120B checkpoint loaded from a Kaggle model dataset.
\nThe ensemble is created at inference time:
\n(seed + attempt_index) ** 2, with seed=42.temperature=1.0 and min_p=0.02.mean_entropy = average per-token Shannon entropy over top logprobs\nweight = 1 / max(mean_entropy, entropy_floor)\nentropy_floor = 1e-9\nThe answer with the largest total weight is submitted. If no valid answer is found, the notebook returns 0.
The most important result was that inference-system reliability mattered more than adding more reasoning machinery. Strong leaderboard submissions consistently used pass@8, short prompts, FP8 KV cache, prefix caching, long context, and a 2x overprovisioned sandbox pool.
\nKey findings from the release history:
\ncontext_tokens=81920 avoided truncation failures observed at 65,536 tokens while preserving enough concurrency for 8 attempts.workers=16 for attempts=8) repeatedly outperformed 1:1 worker/attempt configurations on the public leaderboard, even though the 10-problem local reference subset did not reveal the difference.gpu_memory_utilization=0.99.performance_mode=\"throughput\" was preferred over interactivity because prefix caching reduced repeated prefill cost and generation throughput dominated.The later Mk II, Mk III, and Mk IV notebooks were useful research lines, but they are not the publicly available Kaggle notebook lineage. They tested superior-in-theory approaches such as vLLM upgrades, longer contexts, richer sandbox cleanup, explicit attention configuration, CUDA graph tuning, Marlin MoE backend selection, developer-channel prompt layouts, extra error interception, and XML-style prompts. Despite these improvements in engineering maturity, the follow-up pipelines did not outperform the public no-suffix notebook lineage's 42/50 public leaderboard peak.
\nSeveral explored ideas were not used because they reduced reliability or exceeded the runtime budget: fine-tuning, retrieval-augmented generation, in-context example retrieval, speculative decoding, entropy pruning from larger pass@K runs, sequential refinement, extra DSL/prover tools, developer-channel prompt variants, and eager-execution router/probing experiments.
\nThe closest simplified model that preserves most of the final result is the early Mk I baseline: GPT-OSS-120B with the Python tool, pass@8, prefix caching, 65,536-token context, and frequency-based voting. That baseline reached 40/50 on the public leaderboard, which is close to the selected Version 16 score of 41/50 and the public lineage peak of 42/50.
\nThe most important single model is GPT-OSS-120B. The most important method is not training but reliable inference-time sampling with Python tool use and answer aggregation.
\nA very simple version with only one attempt would be faster but was not the selected approach because public leaderboard variance and hard-problem coverage favored multiple independent attempts.
\nPrediction time is bounded by the Kaggle notebook limit. The final selected notebook is configured for the 5-hour competition environment with:
\n| Runtime Setting | \nValue | \n
|---|---|
| Notebook limit | \n17,400 seconds | \n
| Server startup timeout | \n180 seconds | \n
| Base problem timeout | \n240 seconds | \n
| High problem timeout | \n840 seconds | \n
| Session timeout | \n900 seconds | \n
| Python tool timeout | \n6 seconds | \n
| Problems | \n50 | \n
| Attempts per problem | \n8 | \n
| Sandbox workers | \n16 | \n
Actual runtime is stochastic because problem difficulty, answer agreement, tool calls, and generated token counts vary. The pipeline reserves time for remaining problems and gives surplus time to current hard problems while trying to finish within the competition's 5-hour limit.
\nThe simpler Mk I-style baselines also perform no training. Their prediction runtime is similar or lower, but the selected Version 16 notebook combined tighter time allocation, refactored streaming/tool handling, regex compilation, and prefix caching in a more reliable public-notebook configuration.
\nRelevant references and sources used during development:
\ncode/inference/**/RELEASES.md.The solution can be submitted as a Kaggle notebook solution. The selected final inference notebook is the public no-suffix Mk I notebook:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThe utility notebook used to prepare offline wheels and tiktoken encoding files is:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-utils.ipynb\nThe later Mk II, Mk III, and Mk IV directories document follow-up experiments. They are useful for explaining the development process, but the no-suffix gpt-oss-120b-with-tools lineage is the public notebook lineage, and Version 16 is the selected final submission notebook.
The release notes that document development history are under:
\ncode/inference/gpt-oss-120b-with-tools/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-ii-previews/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-ii/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iii-previews/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iii/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iv/RELEASES.md\nNo trained model file is produced by this repository. The pretrained model is loaded from the Kaggle GPT-OSS-120B model input.
\nThe Kaggle competition notebook environment provides:
\n| Resource | \nValue | \n
|---|---|
| GPU | \n1 x NVIDIA H100 SXM | \n
| GPU memory | \n80 GB HBM3 | \n
| GPU bandwidth | \n3.35 TB/s | \n
| CPU workers | \n26 | \n
| System memory | \n230 GB | \n
| Internet | \nDisabled during competition inference | \n
The selected notebook metadata records:
\n| Platform Item | \nValue | \n
|---|---|
| Environment | \nKaggle Notebook | \n
| Accelerator | \nNVIDIA H100 | \n
| Python | \n3.12.12 | \n
| Kaggle Docker image version | \n31236 | \n
| Internet | \nDisabled for inference notebook | \n
The selected public utility notebook downloads offline wheels for:
\nunsloth\ntrl\nvllm==0.11.2\nopenai_harmony\nThe utility notebook downloads Python 3.12 binary wheels using the CUDA 12.8 PyTorch index and the NVIDIA package index, then packages the wheels and tiktoken encodings into wheels.tar.gz. The selected inference notebook extracts that archive to:
/kaggle/tmp/setup\nand installs from:
\n/kaggle/tmp/setup/wheels\nImportant runtime libraries imported by the selected notebook include vllm, openai, openai_harmony, transformers, jupyter_client, pandas, polars, and Kaggle's kaggle_evaluation.aimo_3_inference_server.
There is no training step.
\nRun the selected Mk I Version 16 Kaggle notebook with the required Kaggle inputs attached:
\n/kaggle/input/gpt-oss-120b/transformers/default/1./kaggle/input/aimo-3-utils/wheels.tar.gz.The notebook:
\n/kaggle/input/aimo-3-utils/wheels.tar.gz to /kaggle/tmp/setup.unsloth, trl, vllm, and openai_harmony from offline wheels.predict() with kaggle_evaluation.aimo_3_inference_server.AIMO3InferenceServer.serve() during competition reruns or run_local_gateway() for local reference execution.The notebook intentionally modifies the runtime environment:
\nkeras, matplotlib, scikit-learn, and tensorflow./kaggle/tmp/setup.8000.vllm_server.log in the working directory.50000.aimo-3-utils wheel archive are attached as Kaggle inputs.KAGGLE_IS_COMPETITION_RERUN.There are no external configuration files required by the final notebook. Runtime configuration is centralized in the CFG class inside the selected notebook.
Important final configuration values include:
\n| Setting | \nValue | \n
|---|---|
| Model | \nGPT-OSS-120B | \n
| Model path | \n/kaggle/input/gpt-oss-120b/transformers/default/1 | \n
| vLLM version | \n0.11.2 | \n
| Reasoning effort | \nHIGH | \n
| KV cache dtype | \nfp8_e4m3 | \n
| dtype | \nauto | \n
| Context length | \n65,536 | \n
| Search tokens | \n32 | \n
| Buffer tokens | \n512 | \n
| Max sequences | \n256 | \n
| GPU memory utilization | \n0.96 | \n
| Prefix caching | \nEnabled | \n
| Async scheduling | \nEnabled | \n
| Disable log stats | \nEnabled | \n
| Temperature | \n1.0 | \n
| min_p | \n0.02 | \n
| top_logprobs | \n5 | \n
| Attempts | \n8 | \n
| Workers | \n16 | \n
| Early stop | \n4 matching answers | \n
| Max turns | \n128 | \n
| Seed | \n42 | \n
| Entropy score floor | \n1e-9 in the inverse-entropy weighting expression | \n
For the Kaggle notebook workflow, exact package installation is handled by the utility dataset. The repository requirements.txt corresponds to the utility notebook used by the selected Mk I Version 16 submission, code/inference/gpt-oss-120b-with-tools/aimo-3-utils.ipynb; it is the expanded pinned package list derived from the wheels downloaded by that utility notebook.
The selected utility notebook downloads binary wheels for Python 3.12 using:
\npip download --dest /kaggle/working/wheels --python-version 3.12 --only-binary=:all: --extra-index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.nvidia.com unsloth trl vllm==0.11.2 openai_harmony\nThe inference notebook then installs offline with:
\npython -m pip install --no-index --find-links /kaggle/tmp/setup/wheels unsloth trl vllm openai_harmony\nThe tiktoken encoding files are bundled into the same archive and loaded from:
\n/kaggle/tmp/setup/tiktoken_encodings\nThe main project directories relevant to this writeup are:
\n.\ncode/\ncode/inference/\ncode/inference/gpt-oss-120b-with-tools/\ncode/inference/gpt-oss-120b-with-tools-mk-ii-previews/\ncode/inference/gpt-oss-120b-with-tools-mk-ii/\ncode/inference/gpt-oss-120b-with-tools-mk-iii-previews/\ncode/inference/gpt-oss-120b-with-tools-mk-iii/\ncode/inference/gpt-oss-120b-with-tools-mk-iv/\ncode/other/gpt-oss-120b-error-logs/\nsandbox/\nFor a formal archive, generate directory_structure.txt from the project root with:
find . -type d > directory_structure.txt\nKaggle's generic documentation template recommends a SETTINGS.json file. This repository uses a notebook-style competition submission instead. All I/O paths and runtime settings are centralized in the notebook's CFG class and in the Kaggle input attachments.
Primary paths:
\n| Purpose | \nPath | \n
|---|---|
| Model weights | \n/kaggle/input/gpt-oss-120b/transformers/default/1 | \n
| Utility archive | \n/kaggle/input/aimo-3-utils/wheels.tar.gz | \n
| Extracted offline wheels | \n/kaggle/tmp/setup/wheels | \n
| Tiktoken encodings | \n/kaggle/tmp/setup/tiktoken_encodings | \n
| Local gateway data | \n/kaggle/input/ai-mathematical-olympiad-progress-prize-3/test.csv | \n
| vLLM server log | \nvllm_server.log | \n
No serialized trained model is produced. The submission uses the pretrained GPT-OSS-120B checkpoint supplied as a Kaggle model input. All competition-specific behavior is implemented in the inference notebook.
\nThe competition entry point is the predict() function in the selected notebook:
code/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThe notebook creates:
\ninference_server = kaggle_evaluation.aimo_3_inference_server.AIMO3InferenceServer(predict)\nThen it uses:
\ninference_server.serve()\nwhen KAGGLE_IS_COMPETITION_RERUN is set. For local reference execution, it uses:
inference_server.run_local_gateway((\"/kaggle/input/ai-mathematical-olympiad-progress-prize-3/test.csv\",))\nThere are no separate prepare_data.py, train.py, or predict.py scripts because the Kaggle submission format is a notebook with a registered inference server callback.
I think you are the real champion on this challenge, respect you!
", "votes": 3, "replies": [] }, { "id": 3461590, "author_name": "rıza temizel", "author_url": "", "post_date": "2026-05-20T22:04:15.470000", "content": "At least for this competition, your work feels like an “Attention Is All You Need” level contribution. Thank you for sharing it.
", "votes": 3, "replies": [ { "id": 3461592, "author_name": "Andreas Bisiadis", "author_url": "", "post_date": "2026-05-20T22:11:50.640000", "content": "\n\nyour work feels like an “Attention Is All You Need” level contribution.
\n
Thank you. That praise means a lot.
", "votes": 1, "replies": [] } ] }, { "id": 3467190, "author_name": "Evan Harbaugh", "author_url": "", "post_date": "2026-06-05T16:34:01.877000", "content": "Excellent writeup! Phenomenal work.
", "votes": 1, "replies": [] }, { "id": 3463840, "author_name": "Navneet", "author_url": "", "post_date": "2026-05-28T04:56:22.743000", "content": "Cool WriteUp on GPT-OSS-120B (with tools) @andreasbis
", "votes": 1, "replies": [] }, { "id": 3464276, "author_name": "Laureano Arcanio", "author_url": "", "post_date": "2026-05-29T11:38:53.467000", "content": "Thanks for this amazing contribution.\nI honestly wish you had won the competition, it would have been totally deserved.\nLooking forward to seeing you win the writeup prize.
", "votes": 2, "replies": [ { "id": 3473101, "author_name": "Simon Frieder", "author_url": "", "post_date": "2026-06-15T19:45:38.947000", "content": "Just FYI, Andreas Bisiadis didn't enter his write-up towards the write-up prize (i.e., he didn't leave a comment here to enter it: https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-3/discussion/689703)
", "votes": 1, "replies": [] } ] }, { "id": 3461921, "author_name": "Durga Kumari", "author_url": "", "post_date": "2026-05-21T18:50:42.673000", "content": "Really impressive engineering work. Especially the focus on inference reliability and tool orchestration instead of overcomplicating the pipeline.
", "votes": 1, "replies": [] }, { "id": 3461678, "author_name": "Kashyapsinh gohil", "author_url": "", "post_date": "2026-05-21T06:10:18.193000", "content": "Thank you so much for this amazing work. It has effectively become the foundation for the entire competition. Almost everyone here has used your approach and adapted it according to their own knowledge. We owe a massive part of our progress with our work to your groundwork. You truly deserve the title of Best contributor in this competition for the sheer impact you’ve had on everyone. Thank you.
", "votes": 2, "replies": [] }, { "id": 3461651, "author_name": "Onur Fidaner", "author_url": "", "post_date": "2026-05-21T03:07:18.877000", "content": "Thank you for putting this writeup together, and of course sharing your code in the first place. Your notebook was the foundation for many of us, including myself. Incredibly efficient pipeline! Seeing the thought process behind some of the design choices is very useful.
", "votes": 2, "replies": [] }, { "id": 3462122, "author_name": "", "author_url": "", "post_date": "2026-05-22T08:09:51", "content": "", "votes": 0, "replies": [] } ], "comment_count": 11, "messages": [ { "id": 3461570, "postDate": "2026-05-20T20:41:43.727Z", "content": "| Field | \nValue | \n
|---|---|
| Competition Name | \nAI Mathematical Olympiad - Progress Prize 3 | \n
| Team Name | \nAndreas Bisiadis | \n
| Private Leaderboard Score | \n41.5 | \n
| Private Leaderboard Place | \n546 | \n
| Public Leaderboard Reference | \n41/50 for the selected Mk I Version 16 notebook; 42/50 peak for the public no-suffix notebook lineage | \n
| Team Member Name | \nAndreas Bisiadis | \n
This submission was developed for Kaggle's AI Mathematical Olympiad - Progress Prize 3 competition. The selected solution is an inference-only notebook based on GPT-OSS-120B with tool-integrated Python reasoning. The notebook selected for the final submission was Mk I Version 16:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThis selected notebook scored 41/50 on the public leaderboard and 41.5 on the private leaderboard. The same public, no-suffix notebook lineage reached a peak public score of 42/50 in Version 11.
\nThis was a solo submission by Andreas Bisiadis. The technical work represented in this repository focused on long-context LLM inference systems, vLLM backend tuning, tool-integrated mathematical reasoning, Jupyter kernel sandbox reliability, prompt minimization, and answer aggregation under Kaggle's fixed H100 runtime constraints.
\nPrimary competition work included:
\ngpt-oss-120b-with-tools, through 22 versions. This is the only lineage intended to correspond to the publicly available Kaggle notebook.vllm==0.17.0, while the selected public Mk I Version 16 notebook used vllm==0.11.2.The final selected solution is an inference-time ensemble built around GPT-OSS-120B served locally with vLLM on a single Kaggle H100 GPU. Each math problem is solved by up to 8 parallel attempts using the Harmony prompt format, ReasoningEffort.HIGH, and a Python tool backed by persistent Jupyter kernels. Candidate integer answers are extracted from final answers using a \\boxed{} parser with a narrow textual fallback for \"final answer is\", then aggregated with inverse-entropy weighting computed from top-logprob token entropy. The selected Mk I Version 16 notebook uses vllm==0.11.2, a 65,536-token context, FP8 E4M3 KV cache, async scheduling, disabled log stats, and prefix caching. No model training or fine-tuning is performed; development effort was concentrated on inference robustness, tool reliability, runtime allocation, and answer selection.
The effective inputs and engineered signals are:
\n\\boxed{} format.math, numpy, and sympy.The main transformations are prompt composition, answer normalization, and entropy-based scoring. Boxed answer extraction accepts comma separators, strips them, converts to an integer, and keeps only values in [0, 99999]. Version 16 also includes a narrow case-insensitive fallback for text matching final answer is <integer>.
No external problem data or retrieval corpus is used. The only external artifacts are allowed Kaggle inputs: the competition data, the GPT-OSS-120B model weights, utility wheels, and local tiktoken encoding files needed for offline execution.
\nNo training, fine-tuning, reinforcement learning, or preference optimization is used in the submitted solution. The model is the pretrained GPT-OSS-120B checkpoint loaded from a Kaggle model dataset.
\nThe ensemble is created at inference time:
\n(seed + attempt_index) ** 2, with seed=42.temperature=1.0 and min_p=0.02.mean_entropy = average per-token Shannon entropy over top logprobs\nweight = 1 / max(mean_entropy, entropy_floor)\nentropy_floor = 1e-9\nThe answer with the largest total weight is submitted. If no valid answer is found, the notebook returns 0.
The most important result was that inference-system reliability mattered more than adding more reasoning machinery. Strong leaderboard submissions consistently used pass@8, short prompts, FP8 KV cache, prefix caching, long context, and a 2x overprovisioned sandbox pool.
\nKey findings from the release history:
\ncontext_tokens=81920 avoided truncation failures observed at 65,536 tokens while preserving enough concurrency for 8 attempts.workers=16 for attempts=8) repeatedly outperformed 1:1 worker/attempt configurations on the public leaderboard, even though the 10-problem local reference subset did not reveal the difference.gpu_memory_utilization=0.99.performance_mode=\"throughput\" was preferred over interactivity because prefix caching reduced repeated prefill cost and generation throughput dominated.The later Mk II, Mk III, and Mk IV notebooks were useful research lines, but they are not the publicly available Kaggle notebook lineage. They tested superior-in-theory approaches such as vLLM upgrades, longer contexts, richer sandbox cleanup, explicit attention configuration, CUDA graph tuning, Marlin MoE backend selection, developer-channel prompt layouts, extra error interception, and XML-style prompts. Despite these improvements in engineering maturity, the follow-up pipelines did not outperform the public no-suffix notebook lineage's 42/50 public leaderboard peak.
\nSeveral explored ideas were not used because they reduced reliability or exceeded the runtime budget: fine-tuning, retrieval-augmented generation, in-context example retrieval, speculative decoding, entropy pruning from larger pass@K runs, sequential refinement, extra DSL/prover tools, developer-channel prompt variants, and eager-execution router/probing experiments.
\nThe closest simplified model that preserves most of the final result is the early Mk I baseline: GPT-OSS-120B with the Python tool, pass@8, prefix caching, 65,536-token context, and frequency-based voting. That baseline reached 40/50 on the public leaderboard, which is close to the selected Version 16 score of 41/50 and the public lineage peak of 42/50.
\nThe most important single model is GPT-OSS-120B. The most important method is not training but reliable inference-time sampling with Python tool use and answer aggregation.
\nA very simple version with only one attempt would be faster but was not the selected approach because public leaderboard variance and hard-problem coverage favored multiple independent attempts.
\nPrediction time is bounded by the Kaggle notebook limit. The final selected notebook is configured for the 5-hour competition environment with:
\n| Runtime Setting | \nValue | \n
|---|---|
| Notebook limit | \n17,400 seconds | \n
| Server startup timeout | \n180 seconds | \n
| Base problem timeout | \n240 seconds | \n
| High problem timeout | \n840 seconds | \n
| Session timeout | \n900 seconds | \n
| Python tool timeout | \n6 seconds | \n
| Problems | \n50 | \n
| Attempts per problem | \n8 | \n
| Sandbox workers | \n16 | \n
Actual runtime is stochastic because problem difficulty, answer agreement, tool calls, and generated token counts vary. The pipeline reserves time for remaining problems and gives surplus time to current hard problems while trying to finish within the competition's 5-hour limit.
\nThe simpler Mk I-style baselines also perform no training. Their prediction runtime is similar or lower, but the selected Version 16 notebook combined tighter time allocation, refactored streaming/tool handling, regex compilation, and prefix caching in a more reliable public-notebook configuration.
\nRelevant references and sources used during development:
\ncode/inference/**/RELEASES.md.The solution can be submitted as a Kaggle notebook solution. The selected final inference notebook is the public no-suffix Mk I notebook:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThe utility notebook used to prepare offline wheels and tiktoken encoding files is:
\ncode/inference/gpt-oss-120b-with-tools/aimo-3-utils.ipynb\nThe later Mk II, Mk III, and Mk IV directories document follow-up experiments. They are useful for explaining the development process, but the no-suffix gpt-oss-120b-with-tools lineage is the public notebook lineage, and Version 16 is the selected final submission notebook.
The release notes that document development history are under:
\ncode/inference/gpt-oss-120b-with-tools/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-ii-previews/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-ii/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iii-previews/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iii/RELEASES.md\ncode/inference/gpt-oss-120b-with-tools-mk-iv/RELEASES.md\nNo trained model file is produced by this repository. The pretrained model is loaded from the Kaggle GPT-OSS-120B model input.
\nThe Kaggle competition notebook environment provides:
\n| Resource | \nValue | \n
|---|---|
| GPU | \n1 x NVIDIA H100 SXM | \n
| GPU memory | \n80 GB HBM3 | \n
| GPU bandwidth | \n3.35 TB/s | \n
| CPU workers | \n26 | \n
| System memory | \n230 GB | \n
| Internet | \nDisabled during competition inference | \n
The selected notebook metadata records:
\n| Platform Item | \nValue | \n
|---|---|
| Environment | \nKaggle Notebook | \n
| Accelerator | \nNVIDIA H100 | \n
| Python | \n3.12.12 | \n
| Kaggle Docker image version | \n31236 | \n
| Internet | \nDisabled for inference notebook | \n
The selected public utility notebook downloads offline wheels for:
\nunsloth\ntrl\nvllm==0.11.2\nopenai_harmony\nThe utility notebook downloads Python 3.12 binary wheels using the CUDA 12.8 PyTorch index and the NVIDIA package index, then packages the wheels and tiktoken encodings into wheels.tar.gz. The selected inference notebook extracts that archive to:
/kaggle/tmp/setup\nand installs from:
\n/kaggle/tmp/setup/wheels\nImportant runtime libraries imported by the selected notebook include vllm, openai, openai_harmony, transformers, jupyter_client, pandas, polars, and Kaggle's kaggle_evaluation.aimo_3_inference_server.
There is no training step.
\nRun the selected Mk I Version 16 Kaggle notebook with the required Kaggle inputs attached:
\n/kaggle/input/gpt-oss-120b/transformers/default/1./kaggle/input/aimo-3-utils/wheels.tar.gz.The notebook:
\n/kaggle/input/aimo-3-utils/wheels.tar.gz to /kaggle/tmp/setup.unsloth, trl, vllm, and openai_harmony from offline wheels.predict() with kaggle_evaluation.aimo_3_inference_server.AIMO3InferenceServer.serve() during competition reruns or run_local_gateway() for local reference execution.The notebook intentionally modifies the runtime environment:
\nkeras, matplotlib, scikit-learn, and tensorflow./kaggle/tmp/setup.8000.vllm_server.log in the working directory.50000.aimo-3-utils wheel archive are attached as Kaggle inputs.KAGGLE_IS_COMPETITION_RERUN.There are no external configuration files required by the final notebook. Runtime configuration is centralized in the CFG class inside the selected notebook.
Important final configuration values include:
\n| Setting | \nValue | \n
|---|---|
| Model | \nGPT-OSS-120B | \n
| Model path | \n/kaggle/input/gpt-oss-120b/transformers/default/1 | \n
| vLLM version | \n0.11.2 | \n
| Reasoning effort | \nHIGH | \n
| KV cache dtype | \nfp8_e4m3 | \n
| dtype | \nauto | \n
| Context length | \n65,536 | \n
| Search tokens | \n32 | \n
| Buffer tokens | \n512 | \n
| Max sequences | \n256 | \n
| GPU memory utilization | \n0.96 | \n
| Prefix caching | \nEnabled | \n
| Async scheduling | \nEnabled | \n
| Disable log stats | \nEnabled | \n
| Temperature | \n1.0 | \n
| min_p | \n0.02 | \n
| top_logprobs | \n5 | \n
| Attempts | \n8 | \n
| Workers | \n16 | \n
| Early stop | \n4 matching answers | \n
| Max turns | \n128 | \n
| Seed | \n42 | \n
| Entropy score floor | \n1e-9 in the inverse-entropy weighting expression | \n
For the Kaggle notebook workflow, exact package installation is handled by the utility dataset. The repository requirements.txt corresponds to the utility notebook used by the selected Mk I Version 16 submission, code/inference/gpt-oss-120b-with-tools/aimo-3-utils.ipynb; it is the expanded pinned package list derived from the wheels downloaded by that utility notebook.
The selected utility notebook downloads binary wheels for Python 3.12 using:
\npip download --dest /kaggle/working/wheels --python-version 3.12 --only-binary=:all: --extra-index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.nvidia.com unsloth trl vllm==0.11.2 openai_harmony\nThe inference notebook then installs offline with:
\npython -m pip install --no-index --find-links /kaggle/tmp/setup/wheels unsloth trl vllm openai_harmony\nThe tiktoken encoding files are bundled into the same archive and loaded from:
\n/kaggle/tmp/setup/tiktoken_encodings\nThe main project directories relevant to this writeup are:
\n.\ncode/\ncode/inference/\ncode/inference/gpt-oss-120b-with-tools/\ncode/inference/gpt-oss-120b-with-tools-mk-ii-previews/\ncode/inference/gpt-oss-120b-with-tools-mk-ii/\ncode/inference/gpt-oss-120b-with-tools-mk-iii-previews/\ncode/inference/gpt-oss-120b-with-tools-mk-iii/\ncode/inference/gpt-oss-120b-with-tools-mk-iv/\ncode/other/gpt-oss-120b-error-logs/\nsandbox/\nFor a formal archive, generate directory_structure.txt from the project root with:
find . -type d > directory_structure.txt\nKaggle's generic documentation template recommends a SETTINGS.json file. This repository uses a notebook-style competition submission instead. All I/O paths and runtime settings are centralized in the notebook's CFG class and in the Kaggle input attachments.
Primary paths:
\n| Purpose | \nPath | \n
|---|---|
| Model weights | \n/kaggle/input/gpt-oss-120b/transformers/default/1 | \n
| Utility archive | \n/kaggle/input/aimo-3-utils/wheels.tar.gz | \n
| Extracted offline wheels | \n/kaggle/tmp/setup/wheels | \n
| Tiktoken encodings | \n/kaggle/tmp/setup/tiktoken_encodings | \n
| Local gateway data | \n/kaggle/input/ai-mathematical-olympiad-progress-prize-3/test.csv | \n
| vLLM server log | \nvllm_server.log | \n
No serialized trained model is produced. The submission uses the pretrained GPT-OSS-120B checkpoint supplied as a Kaggle model input. All competition-specific behavior is implemented in the inference notebook.
\nThe competition entry point is the predict() function in the selected notebook:
code/inference/gpt-oss-120b-with-tools/aimo-3-gpt-oss-120b-with-tools-version-16.ipynb\nThe notebook creates:
\ninference_server = kaggle_evaluation.aimo_3_inference_server.AIMO3InferenceServer(predict)\nThen it uses:
\ninference_server.serve()\nwhen KAGGLE_IS_COMPETITION_RERUN is set. For local reference execution, it uses:
inference_server.run_local_gateway((\"/kaggle/input/ai-mathematical-olympiad-progress-prize-3/test.csv\",))\nThere are no separate prepare_data.py, train.py, or predict.py scripts because the Kaggle submission format is a notebook with a registered inference server callback.
I think you are the real champion on this challenge, respect you!
", "rawMarkdown": "I think you are the real champion on this challenge, respect you!", "votes": 3 }, { "id": 3461590, "postDate": "2026-05-20T22:04:15.470Z", "content": "At least for this competition, your work feels like an “Attention Is All You Need” level contribution. Thank you for sharing it.
", "rawMarkdown": "At least for this competition, your work feels like an “Attention Is All You Need” level contribution. Thank you for sharing it.", "votes": 3, "replies": [ { "id": 3461592, "postDate": "2026-05-20T22:11:50.640Z", "content": "\n\nyour work feels like an “Attention Is All You Need” level contribution.
\n
Thank you. That praise means a lot.
", "rawMarkdown": "> your work feels like an “Attention Is All You Need” level contribution.\n\nThank you. That praise means a lot.", "votes": 1 } ] }, { "id": 3467190, "postDate": "2026-06-05T16:34:01.877Z", "content": "Excellent writeup! Phenomenal work.
", "rawMarkdown": "Excellent writeup! Phenomenal work.", "votes": 1 }, { "id": 3463840, "postDate": "2026-05-28T04:56:22.743Z", "content": "Cool WriteUp on GPT-OSS-120B (with tools) @andreasbis
", "rawMarkdown": "Cool WriteUp on GPT-OSS-120B (with tools) @andreasbis ", "votes": 1 }, { "id": 3464276, "postDate": "2026-05-29T11:38:53.467Z", "content": "Thanks for this amazing contribution.\nI honestly wish you had won the competition, it would have been totally deserved.\nLooking forward to seeing you win the writeup prize.
", "rawMarkdown": "Thanks for this amazing contribution.\nI honestly wish you had won the competition, it would have been totally deserved.\nLooking forward to seeing you win the writeup prize.", "votes": 2, "replies": [ { "id": 3473101, "postDate": "2026-06-15T19:45:38.947Z", "content": "Just FYI, Andreas Bisiadis didn't enter his write-up towards the write-up prize (i.e., he didn't leave a comment here to enter it: https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-3/discussion/689703)
", "rawMarkdown": "Just FYI, Andreas Bisiadis didn't enter his write-up towards the write-up prize (i.e., he didn't leave a comment here to enter it: https://www.kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-3/discussion/689703)", "votes": 1 } ] }, { "id": 3461921, "postDate": "2026-05-21T18:50:42.673Z", "content": "Really impressive engineering work. Especially the focus on inference reliability and tool orchestration instead of overcomplicating the pipeline.
", "rawMarkdown": "Really impressive engineering work. Especially the focus on inference reliability and tool orchestration instead of overcomplicating the pipeline.", "votes": 1 }, { "id": 3461678, "postDate": "2026-05-21T06:10:18.193Z", "content": "Thank you so much for this amazing work. It has effectively become the foundation for the entire competition. Almost everyone here has used your approach and adapted it according to their own knowledge. We owe a massive part of our progress with our work to your groundwork. You truly deserve the title of Best contributor in this competition for the sheer impact you’ve had on everyone. Thank you.
", "rawMarkdown": "Thank you so much for this amazing work. It has effectively become the foundation for the entire competition. Almost everyone here has used your approach and adapted it according to their own knowledge. We owe a massive part of our progress with our work to your groundwork. You truly deserve the title of **Best contributor** in this competition for the sheer impact you’ve had on everyone. Thank you.", "votes": 2 }, { "id": 3461651, "postDate": "2026-05-21T03:07:18.877Z", "content": "Thank you for putting this writeup together, and of course sharing your code in the first place. Your notebook was the foundation for many of us, including myself. Incredibly efficient pipeline! Seeing the thought process behind some of the design choices is very useful.
", "rawMarkdown": "Thank you for putting this writeup together, and of course sharing your code in the first place. Your notebook was the foundation for many of us, including myself. Incredibly efficient pipeline! Seeing the thought process behind some of the design choices is very useful.", "votes": 2 }, { "id": 3462122, "postDate": "2026-05-22T08:09:51Z", "rawMarkdown": "", "isDeleted": true } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T04:26:35Z", "competition_id": 118448, "competition_slug": "ai-mathematical-olympiad-progress-prize-3", "team_id": 14799130, "solution_url": "/competitions/ai-mathematical-olympiad-progress-prize-3/writeups/gpt-oss-120b-with-tools-technical-writeup" }