{
  "id": 519160,
  "title": "Training LLM to generate queries is challenging in this setting because of non-differential processes like Whoosh",
  "url": "/competitions/uspto-explainable-ai/discussion/519160",
  "author_name": "Robert Lim",
  "post_date": "2024-07-10T01:31:55.679000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I think training LLM requires using the results of whoosh's search process to caculate the loss function. However, Whoosh's search process is non-differentiable, which disables the backpropagation.</p>\n<p>Does anyone have a solution to address this problem?</p>",
  "messages": [
    {
      "id": 2914494,
      "postDate": "2024-07-10T01:31:55.680Z",
      "content": "<p>I think training LLM requires using the results of whoosh's search process to caculate the loss function. However, Whoosh's search process is non-differentiable, which disables the backpropagation.</p>\n<p>Does anyone have a solution to address this problem?</p>",
      "rawMarkdown": "I think training LLM requires using the results of whoosh's search process to caculate the loss function. However, Whoosh's search process is non-differentiable, which disables the backpropagation.\n\nDoes anyone have a solution to address this problem?",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2914494": "I think training LLM requires using the results of whoosh's search process to caculate the loss function. However, Whoosh's search process is non-differentiable, which disables the backpropagation.\n\nDoes anyone have a solution to address this problem?"
  }
}