{ "id": 585255, "title": "what’s the lowest CV/LB a single model can achieve?", "url": "/competitions/waveform-inversion/discussion/585255", "author_name": "uctpexx123", "post_date": "2025-06-19T04:40:14.320000", "votes": 3, "comment_count": 16, "views": 0, "content": "

my single convnext model CV result:

\n

CurveFault_A 5.06
\nCurveFault_B 65.20
\nCurveVel_A 9.21
\nCurveVel_B 34.24
\nFlatFault_A 3.91
\nFlatFault_B 22.00
\nFlatVel_A 2.23
\nFlatVel_B 8.66
\nStyle_A 28.39
\nStyle_B 49.48

\n

Val MAE: 22.84 not yet submitted

\n

Just wondering — what’s the lowest CV/LB a single model can achieve?

", "messages": [ { "id": 3227585, "postDate": "2025-06-19T05:28:01.750Z", "content": "

I trained convnext for 3 days to reach cv 15.5 under Bartley's fold, not submit yet but i guess lb of 19.5, so called gold medal line.

", "rawMarkdown": "I trained convnext for 3 days to reach cv 15.5 under Bartley's fold, not submit yet but i guess lb of 19.5, so called gold medal line. ", "votes": 3, "replies": [ { "id": 3227630, "postDate": "2025-06-19T06:26:27.213Z", "content": "

May I ask how many epochs you trained your ConvNeXt model? Given that Bartley's ConvNeXt-Small baseline achieves only around 33+ on the leaderboard, reaching a CV score of 15.5 after three days of training is quite impressive.

", "rawMarkdown": "May I ask how many epochs you trained your ConvNeXt model? Given that Bartley's ConvNeXt-Small baseline achieves only around 33+ on the leaderboard, reaching a CV score of 15.5 after three days of training is quite impressive.", "replies": [ { "id": 3227641, "postDate": "2025-06-19T06:36:38.850Z", "content": "

maybe around 130 epochs, just killed when improvement is not obvious, so can't remember exact number.
\nThe base version has approximately the same performance with caformer.

", "rawMarkdown": "maybe around 130 epochs, just killed when improvement is not obvious, so can't remember exact number. \nThe base version has approximately the same performance with caformer. " } ] }, { "id": 3228239, "postDate": "2025-06-20T01:08:46.623Z", "content": "

May I ask do you use any additional losses to train the model?

", "rawMarkdown": "May I ask do you use any additional losses to train the model?" } ] }, { "id": 3227537, "postDate": "2025-06-19T04:40:14.320Z", "content": "

my single convnext model CV result:

\n

CurveFault_A 5.06
\nCurveFault_B 65.20
\nCurveVel_A 9.21
\nCurveVel_B 34.24
\nFlatFault_A 3.91
\nFlatFault_B 22.00
\nFlatVel_A 2.23
\nFlatVel_B 8.66
\nStyle_A 28.39
\nStyle_B 49.48

\n

Val MAE: 22.84 not yet submitted

\n

Just wondering — what’s the lowest CV/LB a single model can achieve?

", "rawMarkdown": "my single convnext model CV result:\n=========================\nCurveFault_A 5.06\nCurveFault_B 65.20\nCurveVel_A 9.21\nCurveVel_B 34.24\nFlatFault_A 3.91\nFlatFault_B 22.00\nFlatVel_A 2.23\nFlatVel_B 8.66\nStyle_A 28.39\nStyle_B 49.48\nVal MAE: 22.84 not yet submitted\n=========================\n\n\nJust wondering — what’s the lowest CV/LB a single model can achieve?", "votes": 3 }, { "id": 3227556, "postDate": "2025-06-19T05:00:43.120Z", "content": "

maybe 10 (inferred from train loss), but it takes long time to converge.

", "rawMarkdown": "maybe 10 (inferred from train loss), but it takes long time to converge. ", "votes": 1, "replies": [ { "id": 3227560, "postDate": "2025-06-19T05:04:57.787Z", "content": "

How long did it take to train your model?
\nI trained mine for 8 days (archive CV < 40 at 10 epochs), but it started converging very slowly toward the end—only improving by about 0.02–0.03—so I decided to stop it.

", "rawMarkdown": "How long did it take to train your model?\nI trained mine for 8 days (archive CV < 40 at 10 epochs), but it started converging very slowly toward the end—only improving by about 0.02–0.03—so I decided to stop it.", "replies": [ { "id": 3227561, "postDate": "2025-06-19T05:06:13.960Z", "content": "

Currently only submitted 1 day train result.

", "rawMarkdown": "Currently only submitted 1 day train result." }, { "id": 3227564, "postDate": "2025-06-19T05:08:09.470Z", "content": "

@yuanzhezhou What compute are you training with if you don't mind me asking and how many epochs roughly in a day

", "rawMarkdown": "@yuanzhezhou What compute are you training with if you don't mind me asking and how many epochs roughly in a day" }, { "id": 3227570, "postDate": "2025-06-19T05:13:29.717Z", "content": "

@harshitsheoran I use several 5090s for training. I'm trying to train for longer now and I only have very few chances for long training :)

", "rawMarkdown": "@harshitsheoran I use several 5090s for training. I'm trying to train for longer now and I only have very few chances for long training :) ", "votes": 1 }, { "id": 3227574, "postDate": "2025-06-19T05:15:59.010Z", "content": "

@yuanzhezhou Thank you, I have around the same compute as well, My CV says that I can get about 17.5 LB in also a single day of training

", "rawMarkdown": "@yuanzhezhou Thank you, I have around the same compute as well, My CV says that I can get about 17.5 LB in also a single day of training", "votes": 2 }, { "id": 3227576, "postDate": "2025-06-19T05:16:20.220Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3227577, "postDate": "2025-06-19T05:17:55.677Z", "content": "

I guess it's a bad news for everyone else in this competition 😂

", "rawMarkdown": "I guess it's a bad news for everyone else in this competition 😂" }, { "id": 3227584, "postDate": "2025-06-19T05:26:25.810Z", "content": "

@harshitsheoran May I ask which backbone (convnext or caformer) you used?

", "rawMarkdown": "@harshitsheoran May I ask which backbone (convnext or caformer) you used? " }, { "id": 3227918, "postDate": "2025-06-19T13:00:59.970Z", "content": "

@harshitsheoran single day of training -> 17.5 wait for ur solution after competition --> how many GPUs ?

", "rawMarkdown": "@harshitsheoran **single day of training** -> **17.5** wait for ur solution after competition --> how many GPUs ?" }, { "id": 3227957, "postDate": "2025-06-19T13:31:23.133Z", "content": "

Got any leak in the deleted comment? @yuanzhezhou

", "rawMarkdown": "Got any leak in the deleted comment? @yuanzhezhou " } ] } ] }, { "id": 3227554, "postDate": "2025-06-19T05:00:15.383Z", "content": "

Hi, here are the results from my single model CV:

\n

CurveFault_A 5.60
\nCurveFault_B 59.76
\nCurveVel_A 8.99
\nCurveVel_B 34.82
\nFlatFault_A 4.37
\nFlatFault_B 20.38
\nFlatVel_A 2.44
\nFlatVel_B 8.75
\nStyle_A 30.14
\nStyle_B 45.13

\n

Val MAE: 22.04
\nAnd LB is around ~25

", "rawMarkdown": "Hi, here are the results from my single model CV:\n\nCurveFault_A 5.60\nCurveFault_B 59.76\nCurveVel_A 8.99\nCurveVel_B 34.82\nFlatFault_A 4.37\nFlatFault_B 20.38\nFlatVel_A 2.44\nFlatVel_B 8.75\nStyle_A 30.14\nStyle_B 45.13\n\nVal MAE: 22.04\nAnd LB is around ~25 \n", "votes": 1 } ], "comments": [ { "id": 3227585, "author_name": "hyd", "author_url": "", "post_date": "2025-06-19T05:28:01.750000", "content": "

I trained convnext for 3 days to reach cv 15.5 under Bartley's fold, not submit yet but i guess lb of 19.5, so called gold medal line.

", "votes": 3, "replies": [ { "id": 3227630, "author_name": "AnnieGo", "author_url": "", "post_date": "2025-06-19T06:26:27.213000", "content": "

May I ask how many epochs you trained your ConvNeXt model? Given that Bartley's ConvNeXt-Small baseline achieves only around 33+ on the leaderboard, reaching a CV score of 15.5 after three days of training is quite impressive.

", "votes": 0, "replies": [ { "id": 3227641, "author_name": "hyd", "author_url": "", "post_date": "2025-06-19T06:36:38.850000", "content": "

maybe around 130 epochs, just killed when improvement is not obvious, so can't remember exact number.
\nThe base version has approximately the same performance with caformer.

", "votes": 0, "replies": [] } ] }, { "id": 3228239, "author_name": "", "author_url": "", "post_date": "2025-06-20T01:08:46.623000", "content": "

May I ask do you use any additional losses to train the model?

", "votes": 0, "replies": [] } ] }, { "id": 3227556, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2025-06-19T05:00:43.120000", "content": "

maybe 10 (inferred from train loss), but it takes long time to converge.

", "votes": 1, "replies": [ { "id": 3227560, "author_name": "", "author_url": "", "post_date": "2025-06-19T05:04:57.787000", "content": "

How long did it take to train your model?
\nI trained mine for 8 days (archive CV < 40 at 10 epochs), but it started converging very slowly toward the end—only improving by about 0.02–0.03—so I decided to stop it.

", "votes": 0, "replies": [ { "id": 3227561, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2025-06-19T05:06:13.960000", "content": "

Currently only submitted 1 day train result.

", "votes": 0, "replies": [] }, { "id": 3227564, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-19T05:08:09.470000", "content": "

@yuanzhezhou What compute are you training with if you don't mind me asking and how many epochs roughly in a day

", "votes": 0, "replies": [] }, { "id": 3227570, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2025-06-19T05:13:29.717000", "content": "

@harshitsheoran I use several 5090s for training. I'm trying to train for longer now and I only have very few chances for long training :)

", "votes": 1, "replies": [] }, { "id": 3227574, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-19T05:15:59.010000", "content": "

@yuanzhezhou Thank you, I have around the same compute as well, My CV says that I can get about 17.5 LB in also a single day of training

", "votes": 2, "replies": [] }, { "id": 3227576, "author_name": "", "author_url": "", "post_date": "2025-06-19T05:16:20.220000", "content": "", "votes": 0, "replies": [] }, { "id": 3227577, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2025-06-19T05:17:55.677000", "content": "

I guess it's a bad news for everyone else in this competition 😂

", "votes": 0, "replies": [] }, { "id": 3227584, "author_name": "sasha", "author_url": "", "post_date": "2025-06-19T05:26:25.810000", "content": "

@harshitsheoran May I ask which backbone (convnext or caformer) you used?

", "votes": 0, "replies": [] }, { "id": 3227918, "author_name": "SeshuRaju 🧘‍♂️", "author_url": "", "post_date": "2025-06-19T13:00:59.970000", "content": "

@harshitsheoran single day of training -> 17.5 wait for ur solution after competition --> how many GPUs ?

", "votes": 0, "replies": [] }, { "id": 3227957, "author_name": "SeshuRaju 🧘‍♂️", "author_url": "", "post_date": "2025-06-19T13:31:23.133000", "content": "

Got any leak in the deleted comment? @yuanzhezhou

", "votes": 0, "replies": [] } ] } ] }, { "id": 3227554, "author_name": "", "author_url": "", "post_date": "2025-06-19T05:00:15.383000", "content": "

Hi, here are the results from my single model CV:

\n

CurveFault_A 5.60
\nCurveFault_B 59.76
\nCurveVel_A 8.99
\nCurveVel_B 34.82
\nFlatFault_A 4.37
\nFlatFault_B 20.38
\nFlatVel_A 2.44
\nFlatVel_B 8.75
\nStyle_A 30.14
\nStyle_B 45.13

\n

Val MAE: 22.04
\nAnd LB is around ~25

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3227585": "I trained convnext for 3 days to reach cv 15.5 under Bartley's fold, not submit yet but i guess lb of 19.5, so called gold medal line. ", "3227537": "my single convnext model CV result:\n=========================\nCurveFault_A 5.06\nCurveFault_B 65.20\nCurveVel_A 9.21\nCurveVel_B 34.24\nFlatFault_A 3.91\nFlatFault_B 22.00\nFlatVel_A 2.23\nFlatVel_B 8.66\nStyle_A 28.39\nStyle_B 49.48\nVal MAE: 22.84 not yet submitted\n=========================\n\n\nJust wondering — what’s the lowest CV/LB a single model can achieve?", "3227556": "maybe 10 (inferred from train loss), but it takes long time to converge. ", "3227554": "Hi, here are the results from my single model CV:\n\nCurveFault_A 5.60\nCurveFault_B 59.76\nCurveVel_A 8.99\nCurveVel_B 34.82\nFlatFault_A 4.37\nFlatFault_B 20.38\nFlatVel_A 2.44\nFlatVel_B 8.75\nStyle_A 30.14\nStyle_B 45.13\n\nVal MAE: 22.04\nAnd LB is around ~25 \n" } }