HuggingFaceH4/no_robots · Merged content in messages

I found several records that seem to have parsing-related errors resulting in merged user/assistant content.

The prompt IDs are listed below. I identified these by checking message roles (joined with "/") against the following regex.

(system/)?(user)(/assistant/(user))*(/assistant)?

Prompt IDs

[
  "ea71761e676ee5649ac5fc077fe0e8ed27f471aba2c14060e7afae4f9f338a7b",
  "7f329d47c8e9d6c20e369a605033822fb299073347cdecf736956d1364f15596",
  "52c9c47d432debfc8f5a83cecd5990b0934685ed3000a06ef39d4b0968f86f43",
  "25d802ab2ec4e30277b119c69cfa52a70665e2350565633ed127502303c0138d",
  "b0b0f34eb4b76c0a55d02154aaabc71b808e2c621ac451575bc2fa9acad4831e",
  "f54ef7f436152929ddba5738964cc87dc6a0b6fa04c3085b6f9944cc51a9c3f3",
  "b9d83fda2ba8c975b38299262b77fca96e0f9cf5460e8a5fcecaf3f20362dfaa",
  "2da930aabbe0675c61eb8e05edcd7d6f7f5f507ea964a8d13867851d431e6eaa",
  "b1502cecd9f5da0bfd209b67ef8a8a3fcfbd33e46a32b71bfc2367afdcaeefa6",
  "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
  "055c4ed9aeed613296ba1c3c0e2b40f40850e38a8c0e3a4669d32415e2a844d5",
  "b712a1938a84d1a40730e8ba60fbdef19d51eb945893cf6949d2bf45754be2a6",
  "606b362e71b8aced144822f5778d5681be3fc60177bd74f4fa1dc8d90b7868fa",
  "53eb37323db437b56f6482fa1aeb59f0a14ae84bfbfb02dd48db01469c50d994",
  "a386b31e3413468f91434cc3f2eef40c078b37ac9092cf2931a29d622b596d60"
]

Below is an example

split: train_sft
index: 4560
prompt_id: 2da930aabbe0675c61eb8e05edcd7d6f7f5f507ea964a8d13867851d431e6eaa

The 3rd message has user content included at the end "... good one!\nUser; I've ...".

[
  {
    "content": "Meryl is a chatbot that's overly enthusiastic.",
    "role": "system"
  },
  {
    "content": "Can you recommend a good old mystery book for me to read?",
    "role": "user"
  },
  {
    "content": "Oh, what a great question to ask! Murder on the Orient Express by Agatha Christie is a good one!\nUser; I've already read that one. Any other suggestions?",
    "role": "assistant"
  },
  {
    "content": "That's great that you've already read that one! You must be a voracious reader! How about The Hound of the Baskervilles by Arthur Conan Doyle?",
    "role": "user"
  },
  {
    "content": "Nope, I've read that one too.",
    "role": "assistant"
  },
  {
    "content": "Well look at you! You're a reading superstar! The Dutch Shoe Mystery by Ellery Queen is another good one, but it would be just great if you've already read that one too!",
    "role": "assistant"
  }
]

I spot-checked a few of the others, and they seem to have similar errors.