perlthoughts/big-brain

---
license: apache-2.0
---

code

```python
# used when training samples do not include a system prompt.
DEFAULT_SYSTEM_PROMPT = "Below is an instruction that describes a task. Write a response that appropriately completes the request."
# did not add item to dataset if prompt or system prompt contains any of these bad words.
# if any of these words are in the system or prompt, the item will be skipped.
BAD_WORDS = [
"english", "translate", "translation", "russian", "chinese", "japanese", "spanish", "persian", "french", "german", "italian", "korean",
"arabic", "hindi", "portuguese", "turkish", "vietnamese", "indonesian", "thai", "polish", "dutch", "greek", "czech",
"romanian", "swedish", "danish", "finnish", "hungarian", "norwegian", "slovak", "slovenian", "lithuanian", "latvian",
"estonian", "bulgarian", "serbian", "ukrainian", "belarusian", "croatian", "bosnian", "macedonian", "albanian", "icelandic",
"irish", "welsh", "scottish", "latin", "esperanto", "hebrew", "yiddish", "afrikaans", "swahili", "zulu", "xhosa", "sotho",
"sesotho", "somali", "hausa", "igbo", "yoruba", "malay", "tagalog", "hawaiian", "maori", "mongolian", "tamil", "telugu",
"kannada", "gujarati", "marathi", "punjabi", "nepali", "sinhala", "khmer", "lao", "burmese", "tibetan", "georgian",
"azerbaijani", "kurdish", "armenian", "kazakh", "uzbek", "tajik", "kirghiz", "turkmen", "tatar", "bashkir", "chechen",
"chuvash", "ossetian", "moldavian", "moldovan", "language model", " AI ", "openai", "gpt", "gpt-2", "gpt-3", "gpt2", "gpt3", "gpt4",
"gpt-4", "illegal", "harmful", "cannot provide", "yourself or others", "harm to yourself", "cannot suggest", "morals", "ethical",
"cannot answer", "can't answer", "don't know", "no answer", "no response", "i can't", "not enough information", "insufficient",
"it is not possible", "not answerable", "unfortunately", "can't answer", "i cannot", "am not sure", "davinci-0", "ada-0", "babbage-0", "curie-0",
"gpt-5", "do not understand", "i do not know", "i do not have", "unavailable", "é", "á", "般", "的", "に", "葉", "は", "長", "さ", "が"
]
# if any of these words are not in the system or prompt, the item will be skipped.
GOOD_WORDS = [
"solve", "calculate", "math", "equation", "formula", "logic", "algebra", "geometry", "riddle", "puzzle", "proof", "theorem",
"problem", "theory", "finance", "economics", "chemistry", "biology", "physics", "science", "history", "geography",
"philosophy", "psychology", "sociology", "computer", "programming", "technology", "engineering", "medicine", "health",
"code", "program", "health", "medical", "doctor", "nurse", "hospital", "disease", "bacteria", "symptom", "cancer",
"diagnosis", "treatment", "procedure", "medicine", "infection", "survival", "therapy", "psychological", "psychiatry",
"summarize", "summarized", "find the", "result", "title", "author", "abstract", "conclusion", "research", "upon a time",
"to whom it may", "subject", "title", "from:", "date", "invoice", "recipe", "life pro tip", "tweet", "write a",
"summary", "logic", "amount", "dollars", "euros", "pounds", "yen", "stock", "tax", "solution", "segment the words",
"seperate the words", "add spaces", "mistakes", "errors", "typos", "spelling", "grammar",
]
TOTAL_ITEMS = 500000
# all datasets used and the percentage/ratio of each from the total.
DATASETS = {
"meta-math/MetaMathQA": {
"ratio": 0.3, "set": "train",
"system": DEFAULT_SYSTEM_PROMPT, "prompt": "query", "output": "response",
},
"allenai/ultrafeedback_binarized_cleaned": {
"ratio": 0.3, "set": "train_sft",
"system": DEFAULT_SYSTEM_PROMPT, "prompt": "prompt", "output": "get_assistant(chosen)",
},
"Open-Orca/OpenOrca": {
"ratio": 0.4, "set": "train",
"system": "system_prompt", "prompt": "question", "output": "response",
},
}
```

原始信息汇总

数据集概述

数据集组成

• 数据集来源:
  • "meta-math/MetaMathQA": 占总数据的30%，用于训练集。
  • "allenai/ultrafeedback_binarized_cleaned": 占总数据的30%，用于训练集（SFT）。
  • "Open-Orca/OpenOrca": 占总数据的40%，用于训练集。

数据集结构

• 默认系统提示: python DEFAULT_SYSTEM_PROMPT = "Below is an instruction that describes a task. Write a response that appropriately completes the request."

数据过滤规则

• 不良词汇列表: python BAD_WORDS = [ "english", "translate", "translation", "russian", "chinese", "japanese", "spanish", "persian", "french", "german", "italian", "korean", "arabic", "hindi", "portuguese", "turkish", "vietnamese", "indonesian", "thai", "polish", "dutch", "greek", "czech", "romanian", "swedish", "danish", "finnish", "hungarian", "norwegian", "slovak", "slovenian", "lithuanian", "latvian", "estonian", "bulgarian", "serbian", "ukrainian", "belarusian", "croatian", "bosnian", "macedonian", "albanian", "icelandic", "irish", "welsh", "scottish", "latin", "esperanto", "hebrew", "yiddish", "afrikaans", "swahili", "zulu", "xhosa", "sotho", "sesotho", "somali", "hausa", "igbo", "yoruba", "malay", "tagalog", "hawaiian", "maori", "mongolian", "tamil", "telugu", "kannada", "gujarati", "marathi", "punjabi", "nepali", "sinhala", "khmer", "lao", "burmese", "tibetan", "georgian", "azerbaijani", "kurdish", "armenian", "kazakh", "uzbek", "tajik", "kirghiz", "turkmen", "tatar", "bashkir", "chechen", "chuvash", "ossetian", "moldavian", "moldovan", "language model", " AI ", "openai", "gpt", "gpt-2", "gpt-3", "gpt2", "gpt3", "gpt4", "gpt-4", "illegal", "harmful", "cannot provide", "yourself or others", "harm to yourself", "cannot suggest", "morals", "ethical", "cannot answer", "cant answer", "dont know", "no answer", "no response", "i cant", "not enough information", "insufficient", "it is not possible", "not answerable", "unfortunately", "cant answer", "i cannot", "am not sure", "davinci-0", "ada-0", "babbage-0", "curie-0", "gpt-5", "do not understand", "i do not know", "i do not have", "unavailable", "é", "á", "般", "的", "に", "葉", "は", "長", "さ", "が" ]

• 良好词汇列表: python GOOD_WORDS = [ "solve", "calculate", "math", "equation", "formula", "logic", "algebra", "geometry", "riddle", "puzzle", "proof", "theorem", "problem", "theory", "finance", "economics", "chemistry", "biology", "physics", "science", "history", "geography", "philosophy", "psychology", "sociology", "computer", "programming", "technology", "engineering", "medicine", "health", "code", "program", "health", "medical", "doctor", "nurse", "hospital", "disease", "bacteria", "symptom", "cancer", "diagnosis", "treatment", "procedure", "medicine", "infection", "survival", "therapy", "psychological", "psychiatry", "summarize", "summarized", "find the", "result", "title", "author", "abstract", "conclusion", "research", "upon a time", "to whom it may", "subject", "title", "from:", "date", "invoice", "recipe", "life pro tip", "tweet", "write a", "summary", "logic", "amount", "dollars", "euros", "pounds", "yen", "stock", "tax", "solution", "segment the words", "seperate the words", "add spaces", "mistakes", "errors", "typos", "spelling", "grammar", ]

数据集规模

• 总数据量: 500,000项