{
  "id": 683329,
  "title": "Issue regarding training speed 0.06 it/s",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/683329",
  "author_name": "",
  "post_date": "2026-03-20T08:55:55.067350200Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello dear competitors, </p>\n<p>I'm new to Kaggle competitions and I have an issue regarding my finetuning speed : My training speed is currently at 0.06 it/s with the following configuration.</p>\n<p>My hardware is this : I have a Nvidia RTX 5060 Ti with 16 Go of memory (use at 100% during training). I have 32 Go Memory ram.\nI'm using a base model of 400 M parameters that can be load on my machine.</p>\n<p>Some of the parameters I'm using : num_beams is set to 1 for training then more for inference. For the tokenizer : max_length is set to 256 with padding set to longest. per_device_train_batch_size=16.</p>\n<p>With this configuration a training speed of 0.06 it/s or 16 s/it seem really slow. I don't know if this is just normal speed or if something in my configuration or Pipeline is broken, or if there are some Akkadian specificities that I'm missing. If anyone could give me some help that would be awesome !</p>\n<p>Also I'm tring to put max_length lower for training then put it higher for inference but I don't know if that would degrade the model performance.  Also the dataset I created has arround 50 000 data so it would take like 15 h+ to train.</p>\n<p>Thank you !</p>",
  "messages": [
    {
      "id": "3425041",
      "postDate": "03/20/2026 08:55:55",
      "content": "<p>Hello dear competitors, </p>\n<p>I'm new to Kaggle competitions and I have an issue regarding my finetuning speed : My training speed is currently at 0.06 it/s with the following configuration.</p>\n<p>My hardware is this : I have a Nvidia RTX 5060 Ti with 16 Go of memory (use at 100% during training). I have 32 Go Memory ram.\nI'm using a base model of 400 M parameters that can be load on my machine.</p>\n<p>Some of the parameters I'm using : num_beams is set to 1 for training then more for inference. For the tokenizer : max_length is set to 256 with padding set to longest. per_device_train_batch_size=16.</p>\n<p>With this configuration a training speed of 0.06 it/s or 16 s/it seem really slow. I don't know if this is just normal speed or if something in my configuration or Pipeline is broken, or if there are some Akkadian specificities that I'm missing. If anyone could give me some help that would be awesome !</p>\n<p>Also I'm tring to put max_length lower for training then put it higher for inference but I don't know if that would degrade the model performance.  Also the dataset I created has arround 50 000 data so it would take like 15 h+ to train.</p>\n<p>Thank you !</p>",
      "rawMarkdown": "Hello dear competitors, \n\nI'm new to Kaggle competitions and I have an issue regarding my finetuning speed : My training speed is currently at 0.06 it/s with the following configuration.\n\nMy hardware is this : I have a Nvidia RTX 5060 Ti with 16 Go of memory (use at 100% during training). I have 32 Go Memory ram.\nI'm using a base model of 400 M parameters that can be load on my machine.\n\nSome of the parameters I'm using : num_beams is set to 1 for training then more for inference. For the tokenizer : max_length is set to 256 with padding set to longest. per_device_train_batch_size=16.\n\nWith this configuration a training speed of 0.06 it/s or 16 s/it seem really slow. I don't know if this is just normal speed or if something in my configuration or Pipeline is broken, or if there are some Akkadian specificities that I'm missing. If anyone could give me some help that would be awesome !\n\nAlso I'm tring to put max_length lower for training then put it higher for inference but I don't know if that would degrade the model performance.  Also the dataset I created has arround 50 000 data so it would take like 15 h+ to train.\n\nThank you !",
      "votes": null
    },
    {
      "id": "3425093",
      "postDate": "03/20/2026 10:39:08",
      "content": "<p>This is a real headache, all the time inverted in this competence was more for data cleaning and \"data fit\" than other issue. I'm using a \"curriculum learning\" to apply LoRA like a waterfall. But the time is the same or maybe more.</p>",
      "rawMarkdown": "This is a real headache, all the time inverted in this competence was more for data cleaning and \"data fit\" than other issue. I'm using a \"curriculum learning\" to apply LoRA like a waterfall. But the time is the same or maybe more.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3425093,
      "author_name": "davidochoacorrales",
      "author_url": "",
      "post_date": "03/20/2026 10:39:08",
      "content": "<p>This is a real headache, all the time inverted in this competence was more for data cleaning and \"data fit\" than other issue. I'm using a \"curriculum learning\" to apply LoRA like a waterfall. But the time is the same or maybe more.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3425041": "Hello dear competitors, \n\nI'm new to Kaggle competitions and I have an issue regarding my finetuning speed : My training speed is currently at 0.06 it/s with the following configuration.\n\nMy hardware is this : I have a Nvidia RTX 5060 Ti with 16 Go of memory (use at 100% during training). I have 32 Go Memory ram.\nI'm using a base model of 400 M parameters that can be load on my machine.\n\nSome of the parameters I'm using : num_beams is set to 1 for training then more for inference. For the tokenizer : max_length is set to 256 with padding set to longest. per_device_train_batch_size=16.\n\nWith this configuration a training speed of 0.06 it/s or 16 s/it seem really slow. I don't know if this is just normal speed or if something in my configuration or Pipeline is broken, or if there are some Akkadian specificities that I'm missing. If anyone could give me some help that would be awesome !\n\nAlso I'm tring to put max_length lower for training then put it higher for inference but I don't know if that would degrade the model performance.  Also the dataset I created has arround 50 000 data so it would take like 15 h+ to train.\n\nThank you !",
    "3425093": "This is a real headache, all the time inverted in this competence was more for data cleaning and \"data fit\" than other issue. I'm using a \"curriculum learning\" to apply LoRA like a waterfall. But the time is the same or maybe more."
  },
  "source": "meta"
}