{ "id": 700269, "title": "3rd Place Solution:Codex-YOLO,Public Annotation Priors,Row-Level Feedback Gating", "content": "

Hi everyone,

\n

I am sharing my 3rd-place private leaderboard solution for FathomNet 2026.

\n

Repository:

\n

https://github.com/dkozlov/fathomnet-2026-3rd-place

\n

This archive includes source code, selected model checkpoints, final submissions, validation scripts, sample inference commands, and a writeup/reproduction guide.

\n

Important attribution note: this 3rd-place solution was completely achieved by Codex GPT-5.5 using codex-yolo in /loop mode:

\n

https://github.com/codex-yolo/codex-yolo

\n

## Final submissions

\n

Final private-best submission:

\n\n

What mattered most

\n

The final gain came less from adding a large number of detector predictions and more from selecting a very small number of reliable score changes.

\n

The private-best CSV can be regenerated from the archived lineage base plus 15 final score edits using the included src/apply_score_changes.py script. The repository also includes\n commands to load the bundled weights and run sample inference on public sample images.

\n

Thanks to the organizers for the competition.

", "comments": [ { "id": 3463636, "author_name": "Dmitry Kozlov", "author_url": "", "post_date": "2026-05-27T14:35:00.760000", "content": "

In addition to https://github.com/codex-yolo/codex-yolo and https://github.com/claude-yolo/claude-yolo I am also glad to present https://mobile-ssh.github.io - A focused SSH client for Android with multi-session terminals, private key authentication, SFTP file transfer, local port forwarding, and tmux-aware controls built for phone and tablet use. If you have friends who'd like to join the beta, please ask them to drop me a line at mobile.ssh.info@gmail.com and I'll add them to the tester list.

", "votes": 0, "replies": [] }, { "id": 3463116, "author_name": "Dmitry Kozlov", "author_url": "", "post_date": "2026-05-25T23:13:00.863000", "content": "

Hi @danielariaslopez, I used local environment with nvidia/cuda:13.1.2-cudnn-devel-ubuntu24.04 container outside of Kaggle and Google Colab.

", "votes": 0, "replies": [] }, { "id": 3463091, "author_name": "Daniel Andrés Arias López", "author_url": "", "post_date": "2026-05-25T20:39:40.690000", "content": "

Hello there! I'd like to know what programming environment you used (e.g., Kaggle, Google Colab, locally on your PC, etc.)

", "votes": 0, "replies": [] }, { "id": 3462607, "author_name": "Dmitry Kozlov", "author_url": "", "post_date": "2026-05-24T06:12:59.393000", "content": "

Hi Rob, thanks for your questions!

\n
\n

How much guidance did you give codex-yolo to get a start on building a solution?

\n
\n

I provided an absolute minimum of hints, codex did its own research.

\n
\n

Did you have an idea of how you wanted it to approach the problem or did you just provide the resources from Kaggle and let it rip?

\n
\n

I had no idea how to approach this task: I hadn't even read the problem description or reviewed the data. I pointed Codex to a Kaggle competition and repeated the exact same prompt every 30 minutes.

\n
\n

Did codex perform EDA to understand the data itself?

\n
\n

I think it did some kind of EDA with grep and adhoc python scripts

\n
\n

How often did you review the solutions it was coming up with?

\n
\n

I didn't review the solutions, but I periodically monitored submissions and if there was no submissions and progress during the day then I was trying to update the prompt to improve the speed/quality

\n
\n

Did you give much corrective feedback?

\n
\n

I provided very little corrective feedback.

\n
\n

I noticed on your github page that you have your \"awesome-knowledge-distillation\" repo (thanks, by the way! starred now).

\n
\n

Thanks for starring https://github.com/dkozlov/awesome-knowledge-distillation (I just updated it with codex-yolo)

\n
\n

I'd imagine having a resource like this for codex to reference and use could be super-helpful! I'm assuming you made use of a that knowledge base or something similar?

\n
\n

No, I did not use any Codex or Claude skills, nor did I use the agent.md files. The less context you preload for Codex or Claude, the more context they can utilize to solve Kaggle tasks.

", "votes": 0, "replies": [] }, { "id": 3460063, "author_name": "Rob Brock", "author_url": "", "post_date": "2026-05-18T15:41:18.737000", "content": "

Awesome! Thanks for sharing.

\n

I'm curious on what your general workflow was for using codex/codex-yolo, and I've got lots of questions!

\n\n

I'm sure I could come up with a lot more questions, but I'll leave it at that for now.

\n

Thanks again for providing your repo/solution. Very interesting stuff and I feel like there is a lot to learn here about utilizing agentic coding efficiently!

\n", "votes": 0, "replies": [] } ], "comment_count": 5, "messages": [ { "id": 3458946, "postDate": "2026-05-17T02:49:10.890Z", "content": "

Hi everyone,

\n

I am sharing my 3rd-place private leaderboard solution for FathomNet 2026.

\n

Repository:

\n

https://github.com/dkozlov/fathomnet-2026-3rd-place

\n

This archive includes source code, selected model checkpoints, final submissions, validation scripts, sample inference commands, and a writeup/reproduction guide.

\n

Important attribution note: this 3rd-place solution was completely achieved by Codex GPT-5.5 using codex-yolo in /loop mode:

\n

https://github.com/codex-yolo/codex-yolo

\n

## Final submissions

\n

Final private-best submission:

\n\n

What mattered most

\n

The final gain came less from adding a large number of detector predictions and more from selecting a very small number of reliable score changes.

\n

The private-best CSV can be regenerated from the archived lineage base plus 15 final score edits using the included src/apply_score_changes.py script. The repository also includes\n commands to load the bundled weights and run sample inference on public sample images.

\n

Thanks to the organizers for the competition.

", "rawMarkdown": "Hi everyone, \n\n I am sharing my 3rd-place private leaderboard solution for FathomNet 2026.\n\n Repository: \n\n https://github.com/dkozlov/fathomnet-2026-3rd-place\n \n This archive includes source code, selected model checkpoints, final submissions, validation scripts, sample inference commands, and a writeup/reproduction guide.\n \n Important attribution note: this 3rd-place solution was completely achieved by Codex GPT-5.5 using `codex-yolo` in `/loop` mode:\n\n https://github.com/codex-yolo/codex-yolo\n\n ## Final submissions\n\n Final private-best submission:\n\n - Private LB: `0.2962`\n - Public LB: `0.2801`\n - File: `fathomnet_02802_public_union_plus_sourcefold993_s995.csv`\n\n Earlier public-UUID/noisy-extension candidate:\n\n - Private LB: `0.2978`\n - Public LB: `0.2799`\n - File: `fathomnet_public_uuid_annotations_all_v2_callistephanus_demote_noisyext_hirudinea_s030_clavularia_yoloonly_s030_multiobs_boost_plus_thr325_s005_nms075.csv`\n\n ## High-level approach\n\n The final solution was not a single-model submission. It was a conservative row-level merge built from:\n\n - YOLO-family detectors trained at high resolution\n - pseudo-label and public/external annotation support\n - RT-DETR support models\n - public UUID annotation priors\n - taxon-specific score calibration and demotion for noisy classes\n - row-level public-feedback gating\n - conservative final score-only edits\n\n The strongest pattern was to use model ensembles to generate/support candidate rows, then use historical public feedback and focused audits to decide which rows were safe to trust.\n\n## What mattered most\n\n The final gain came less from adding a large number of detector predictions and more from selecting a very small number of reliable score changes.\n\n The private-best CSV can be regenerated from the archived lineage base plus 15 final score edits using the included `src/apply_score_changes.py` script. The repository also includes\n commands to load the bundled weights and run sample inference on public sample images.\n\n Thanks to the organizers for the competition.", "votes": 7 }, { "id": 3463636, "postDate": "2026-05-27T14:35:00.760Z", "content": "

In addition to https://github.com/codex-yolo/codex-yolo and https://github.com/claude-yolo/claude-yolo I am also glad to present https://mobile-ssh.github.io - A focused SSH client for Android with multi-session terminals, private key authentication, SFTP file transfer, local port forwarding, and tmux-aware controls built for phone and tablet use. If you have friends who'd like to join the beta, please ask them to drop me a line at mobile.ssh.info@gmail.com and I'll add them to the tester list.

", "rawMarkdown": "In addition to [https://github.com/codex-yolo/codex-yolo](https://github.com/codex-yolo/codex-yolo) and [https://github.com/claude-yolo/claude-yolo](https://github.com/claude-yolo/claude-yolo) I am also glad to present [https://mobile-ssh.github.io](https://mobile-ssh.github.io) - A focused SSH client for Android with multi-session terminals, private key authentication, SFTP file transfer, local port forwarding, and tmux-aware controls built for phone and tablet use. If you have friends who'd like to join the beta, please ask them to drop me a line at mobile.ssh.info@gmail.com and I'll add them to the tester list." }, { "id": 3463116, "postDate": "2026-05-25T23:13:00.863Z", "content": "

Hi @danielariaslopez, I used local environment with nvidia/cuda:13.1.2-cudnn-devel-ubuntu24.04 container outside of Kaggle and Google Colab.

", "rawMarkdown": "Hi @danielariaslopez, I used local environment with `nvidia/cuda:13.1.2-cudnn-devel-ubuntu24.04` container outside of Kaggle and Google Colab." }, { "id": 3463091, "postDate": "2026-05-25T20:39:40.690Z", "content": "

Hello there! I'd like to know what programming environment you used (e.g., Kaggle, Google Colab, locally on your PC, etc.)

", "rawMarkdown": "Hello there! I'd like to know what programming environment you used (e.g., Kaggle, Google Colab, locally on your PC, etc.)" }, { "id": 3462607, "postDate": "2026-05-24T06:12:59.393Z", "content": "

Hi Rob, thanks for your questions!

\n
\n

How much guidance did you give codex-yolo to get a start on building a solution?

\n
\n

I provided an absolute minimum of hints, codex did its own research.

\n
\n

Did you have an idea of how you wanted it to approach the problem or did you just provide the resources from Kaggle and let it rip?

\n
\n

I had no idea how to approach this task: I hadn't even read the problem description or reviewed the data. I pointed Codex to a Kaggle competition and repeated the exact same prompt every 30 minutes.

\n
\n

Did codex perform EDA to understand the data itself?

\n
\n

I think it did some kind of EDA with grep and adhoc python scripts

\n
\n

How often did you review the solutions it was coming up with?

\n
\n

I didn't review the solutions, but I periodically monitored submissions and if there was no submissions and progress during the day then I was trying to update the prompt to improve the speed/quality

\n
\n

Did you give much corrective feedback?

\n
\n

I provided very little corrective feedback.

\n
\n

I noticed on your github page that you have your \"awesome-knowledge-distillation\" repo (thanks, by the way! starred now).

\n
\n

Thanks for starring https://github.com/dkozlov/awesome-knowledge-distillation (I just updated it with codex-yolo)

\n
\n

I'd imagine having a resource like this for codex to reference and use could be super-helpful! I'm assuming you made use of a that knowledge base or something similar?

\n
\n

No, I did not use any Codex or Claude skills, nor did I use the agent.md files. The less context you preload for Codex or Claude, the more context they can utilize to solve Kaggle tasks.

", "rawMarkdown": "Hi Rob, thanks for your questions!\n\n>How much guidance did you give codex-yolo to get a start on building a solution?\n\nI provided an absolute minimum of hints, codex did its own research.\n\n> Did you have an idea of how you wanted it to approach the problem or did you just provide the resources from Kaggle and let it rip?\n\nI had no idea how to approach this task: I hadn't even read the problem description or reviewed the data. I pointed Codex to a Kaggle competition and repeated the exact same prompt every 30 minutes.\n\n> Did codex perform EDA to understand the data itself?\n\nI think it did some kind of EDA with grep and adhoc python scripts\n\n> How often did you review the solutions it was coming up with?\n\nI didn't review the solutions, but I periodically monitored submissions and if there was no submissions and progress during the day then I was trying to update the prompt to improve the speed/quality\n\n> Did you give much corrective feedback?\n\nI provided very little corrective feedback.\n\n> I noticed on your github page that you have your \"awesome-knowledge-distillation\" repo (thanks, by the way! starred now).\n\nThanks for starring https://github.com/dkozlov/awesome-knowledge-distillation (I just updated it with codex-yolo)\n\n> I'd imagine having a resource like this for codex to reference and use could be super-helpful! I'm assuming you made use of a that knowledge base or something similar?\n\nNo, I did not use any Codex or Claude skills, nor did I use the `agent.md` files. The less context you preload for Codex or Claude, the more context they can utilize to solve Kaggle tasks. " }, { "id": 3460063, "postDate": "2026-05-18T15:41:18.737Z", "content": "

Awesome! Thanks for sharing.

\n

I'm curious on what your general workflow was for using codex/codex-yolo, and I've got lots of questions!

\n\n

I'm sure I could come up with a lot more questions, but I'll leave it at that for now.

\n

Thanks again for providing your repo/solution. Very interesting stuff and I feel like there is a lot to learn here about utilizing agentic coding efficiently!

\n", "rawMarkdown": "Awesome! Thanks for sharing.\n\nI'm curious on what your general workflow was for using codex/codex-yolo, and I've got lots of questions!\n\n- How much guidance did you give codex-yolo to get a start on building a solution?\n- Did you have an idea of how you wanted it to approach the problem or did you just provide the resources from Kaggle and let it rip?\n- Did codex perform EDA to understand the data itself?\n- How often did you review the solutions it was coming up with?\n- Did you give much corrective feedback?\n- I noticed on your github page that you have your \"awesome-knowledge-distillation\" repo (thanks, by the way! starred now). I'd imagine having a resource like this for codex to reference and use could be super-helpful! I'm assuming you made use of a that knowledge base or something similar?\n\nI'm sure I could come up with a lot more questions, but I'll leave it at that for now.\n\nThanks again for providing your repo/solution. Very interesting stuff and I feel like there is a lot to learn here about utilizing agentic coding efficiently!\n\n- Rob" } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T20:09:10Z", "competition_id": 125799, "competition_slug": "fathomnet-2026", "team_id": 15655237, "solution_url": "/competitions/fathomnet-2026/writeups/fathomnet-2026-3rd-place" }