Finetune

[ZhangScream]

Hi,i finetune MGM-2B on coco, but i got the warning that:
{'loss': 6.9221, 'grad_norm': tensor(18.7422, device='cuda:0', dtype=torch.float64), 'learning_rate': 9.203084832904885e-06, 'epoch': 0.01} 1%|██▌ | 179/12941 [19:58<23:33:39, 6.65s/it]WARNING: tokenization mismatch: 330 vs. 333. (ignored) WARNING: tokenization mismatch: 309 vs. 312. (ignored) WARNING: tokenization mismatch: 438 vs. 446. (ignored) WARNING: tokenization mismatch: 80 vs. 82. (ignored) WARNING: tokenization mismatch: 388 vs. 395. (ignored) WARNING: tokenization mismatch: 84 vs. 86. (ignored) WARNING: tokenization mismatch: 222 vs. 226. (ignored) WARNING: tokenization mismatch: 207 vs. 211. (ignored) WARNING: tokenization mismatch: 86 vs. 88. (ignored) WARNING: tokenization mismatch: 140 vs. 147. (ignored) WARNING: tokenization mismatch: 155 vs. 163. (ignored) WARNING: tokenization mismatch: 283 vs. 288. (ignored) WARNING: tokenization mismatch: 545 vs. 549. (ignored) WARNING: tokenization mismatch: 543 vs. 546. (ignored)
Is there any problem with the tokenizer_config or minigemini_instruction.json?

[HongLouyemeng]

Hi,i finetune MGM-2B on coco, but i got the warning that: {'loss': 6.9221, 'grad_norm': tensor(18.7422, device='cuda:0', dtype=torch.float64), 'learning_rate': 9.203084832904885e-06, 'epoch': 0.01} 1%|██▌ | 179/12941 [19:58<23:33:39, 6.65s/it]WARNING: tokenization mismatch: 330 vs. 333. (ignored) WARNING: tokenization mismatch: 309 vs. 312. (ignored) WARNING: tokenization mismatch: 438 vs. 446. (ignored) WARNING: tokenization mismatch: 80 vs. 82. (ignored) WARNING: tokenization mismatch: 388 vs. 395. (ignored) WARNING: tokenization mismatch: 84 vs. 86. (ignored) WARNING: tokenization mismatch: 222 vs. 226. (ignored) WARNING: tokenization mismatch: 207 vs. 211. (ignored) WARNING: tokenization mismatch: 86 vs. 88. (ignored) WARNING: tokenization mismatch: 140 vs. 147. (ignored) WARNING: tokenization mismatch: 155 vs. 163. (ignored) WARNING: tokenization mismatch: 283 vs. 288. (ignored) WARNING: tokenization mismatch: 545 vs. 549. (ignored) WARNING: tokenization mismatch: 543 vs. 546. (ignored) Is there any problem with the tokenizer_config or minigemini_instruction.json?

同样的问题，我在自定义数据集赏出现了tokenization mismatch:，loss为0 QAQ

[HongLouyemeng]

修改model vision就行，但训练过程出现：Error in loading 1833, retrying...

[ZhangScream]

你这个model vision指的是--version v1 吗？

[HongLouyemeng]

你这个model vision指的是--version v1 吗？

yes

[ZhangScream]

你这个model vision指的是--version v1 吗？

yes

thanks！改了之后也是同样的问题

[ZhangScream]

修改model vision就行，但训练过程出现：Error in loading 1833, retrying...

这个 version我改成gemma后好了，你报错可能是你的json文件里有的图片地址不存在，我是过滤了他的instruction文件，只保留了coco/train2017这个地址的，不会报错
{'loss': 17.621, 'grad_norm': 1097.9221859404172, 'learning_rate': 5.141388174807198e-08, 'epoch': 0.0} {'loss': 19.8897, 'grad_norm': 1191.090334852173, 'learning_rate': 1.0282776349614396e-07, 'epoch': 0.0} {'loss': 17.617, 'grad_norm': 1020.9409156811793, 'learning_rate': 1.5424164524421595e-07, 'epoch': 0.0} {'loss': 18.9953, 'grad_norm': 1134.2098670413347, 'learning_rate': 2.0565552699228793e-07, 'epoch': 0.0} {'loss': 18.5601, 'grad_norm': 1084.7206032578688, 'learning_rate': 2.5706940874035993e-07, 'epoch': 0.0} {'loss': 17.2726, 'grad_norm': 1040.5268230990873, 'learning_rate': 3.084832904884319e-07, 'epoch': 0.0} {'loss': 18.7021, 'grad_norm': 1164.2017706625636, 'learning_rate': 3.598971722365039e-07, 'epoch': 0.0} {'loss': 18.1496, 'grad_norm': 1044.5027659595357, 'learning_rate': 4.1131105398457585e-07, 'epoch': 0.0} {'loss': 15.6641, 'grad_norm': 1037.2660428641016, 'learning_rate': 4.6272493573264783e-07, 'epoch': 0.0} {'loss': 18.7966, 'grad_norm': 1129.7304014927308, 'learning_rate': 5.141388174807199e-07, 'epoch': 0.0} {'loss': 17.0625, 'grad_norm': 1024.579069798291, 'learning_rate': 5.655526992287918e-07, 'epoch': 0.0} {'loss': 14.5877, 'grad_norm': 830.1190842498937, 'learning_rate': 6.169665809768638e-07, 'epoch': 0.0} {'loss': 15.3645, 'grad_norm': 843.60380558662, 'learning_rate': 6.683804627249357e-07, 'epoch': 0.0} {'loss': 12.0782, 'grad_norm': 325.8663246509485, 'learning_rate': 7.197943444730078e-07, 'epoch': 0.0} {'loss': 11.7363, 'grad_norm': 385.4385170199879, 'learning_rate': 7.712082262210797e-07, 'epoch': 0.0} {'loss': 11.5988, 'grad_norm': 380.7032156615642, 'learning_rate': 8.226221079691517e-07, 'epoch': 0.0} {'loss': 10.91, 'grad_norm': 260.234503595527, 'learning_rate': 8.740359897172238e-07, 'epoch': 0.0} {'loss': 11.4197, 'grad_norm': 353.49476985450883, 'learning_rate': 9.254498714652957e-07, 'epoch': 0.0} {'loss': 11.053, 'grad_norm': 326.06150885488125, 'learning_rate': 9.768637532133676e-07, 'epoch': 0.0} {'loss': 9.7236, 'grad_norm': 144.73662083460025, 'learning_rate': 1.0282776349614397e-06, 'epoch': 0.0} {'loss': 10.4556, 'grad_norm': 275.81074270511215, 'learning_rate': 1.0796915167095116e-06, 'epoch': 0.0} {'loss': 10.1932, 'grad_norm': 258.5231596765796, 'learning_rate': 1.1311053984575837e-06, 'epoch': 0.0}

[HongLouyemeng]

修改model vision就行，但训练过程出现：Error in loading 1833, retrying...

这个 version我改成gemma后好了，你报错可能是你的json文件里有的图片地址不存在，我是过滤了他的instruction文件，只保留了coco/train2017这个地址的，不会报错 {'loss': 17.621, 'grad_norm': 1097.9221859404172, 'learning_rate': 5.141388174807198e-08, 'epoch': 0.0} {'loss': 19.8897, 'grad_norm': 1191.090334852173, 'learning_rate': 1.0282776349614396e-07, 'epoch': 0.0} {'loss': 17.617, 'grad_norm': 1020.9409156811793, 'learning_rate': 1.5424164524421595e-07, 'epoch': 0.0} {'loss': 18.9953, 'grad_norm': 1134.2098670413347, 'learning_rate': 2.0565552699228793e-07, 'epoch': 0.0} {'loss': 18.5601, 'grad_norm': 1084.7206032578688, 'learning_rate': 2.5706940874035993e-07, 'epoch': 0.0} {'loss': 17.2726, 'grad_norm': 1040.5268230990873, 'learning_rate': 3.084832904884319e-07, 'epoch': 0.0} {'loss': 18.7021, 'grad_norm': 1164.2017706625636, 'learning_rate': 3.598971722365039e-07, 'epoch': 0.0} {'loss': 18.1496, 'grad_norm': 1044.5027659595357, 'learning_rate': 4.1131105398457585e-07, 'epoch': 0.0} {'loss': 15.6641, 'grad_norm': 1037.2660428641016, 'learning_rate': 4.6272493573264783e-07, 'epoch': 0.0} {'loss': 18.7966, 'grad_norm': 1129.7304014927308, 'learning_rate': 5.141388174807199e-07, 'epoch': 0.0} {'loss': 17.0625, 'grad_norm': 1024.579069798291, 'learning_rate': 5.655526992287918e-07, 'epoch': 0.0} {'loss': 14.5877, 'grad_norm': 830.1190842498937, 'learning_rate': 6.169665809768638e-07, 'epoch': 0.0} {'loss': 15.3645, 'grad_norm': 843.60380558662, 'learning_rate': 6.683804627249357e-07, 'epoch': 0.0} {'loss': 12.0782, 'grad_norm': 325.8663246509485, 'learning_rate': 7.197943444730078e-07, 'epoch': 0.0} {'loss': 11.7363, 'grad_norm': 385.4385170199879, 'learning_rate': 7.712082262210797e-07, 'epoch': 0.0} {'loss': 11.5988, 'grad_norm': 380.7032156615642, 'learning_rate': 8.226221079691517e-07, 'epoch': 0.0} {'loss': 10.91, 'grad_norm': 260.234503595527, 'learning_rate': 8.740359897172238e-07, 'epoch': 0.0} {'loss': 11.4197, 'grad_norm': 353.49476985450883, 'learning_rate': 9.254498714652957e-07, 'epoch': 0.0} {'loss': 11.053, 'grad_norm': 326.06150885488125, 'learning_rate': 9.768637532133676e-07, 'epoch': 0.0} {'loss': 9.7236, 'grad_norm': 144.73662083460025, 'learning_rate': 1.0282776349614397e-06, 'epoch': 0.0} {'loss': 10.4556, 'grad_norm': 275.81074270511215, 'learning_rate': 1.0796915167095116e-06, 'epoch': 0.0} {'loss': 10.1932, 'grad_norm': 258.5231596765796, 'learning_rate': 1.1311053984575837e-06, 'epoch': 0.0}

好的 谢谢qaq