BramVanroy/dutch-instruction-datasets

Dutch instruction dataset creation

In this repository scripts are provided to build your own instruction dataset through OpenAI services. We specifically make use of Azure services.

Usage

If you use the Azure services in the following scripts, you will need to specify a credentials file. This file should have the following structure, where each key is a "profile", like "gpt-4". In the examples below, this has been saved to a file called .credentials.json.

{
    "gpt-4": {
        "endpoint": "https://abc.openai.azure.com/",
        "api_key": "[secret-key]",
        "api_version": "2023-07-01-preview",
        "deployment_name": "deployment-name1"
    },
    "gpt-35-turbo": {
        "endpoint": "https://def.openai.azure.com/",
        "api_key": "[secret-key]",
        "api_version": "2023-07-01-preview",
        "deployment_name": "deployment-name2"
    }
}

For all commands a --help option is available with more explanations about all the arguments.

interactive-query

Launch an interactive query session. This will allow you to query the OpenAI API and "talk" to the model. This implementation is not very smart, and will not do any smart length filtering when you exceed the context window. So do not use it for extended conversations.

It supports both Azure services and Hugging Face models.

Example usage Azure with the gpt-35-turbo profile:

interactive-query azure .credentials.json gpt-35-turbo

Example usage Hugging Face with the BramVanroy/Llama-2-13b-chat-dutch model (transformers must be installed, and for many options w.r.t. quantization you will also need accelerate and bitsandbytes):

interactive-query huggingface BramVanroy/Llama-2-13b-chat-dutch --load-in-8bit

translate-hf

Most of the time we want to start with translating system message and/or user messages, and then "answer" those later on in a next step. translate-hf is the entry point to translate specific columns and splits of any dataset on the Hugging Face hub. It will save the translated dataset to a temporary location, and then upload it to the hub.

It should be relatively robust as it saves intermediate results and can simply restart where it left off.

Example usage:

translate-hf HuggingFaceH4/ultrachat_200k data/ultrachat_200k/ultrachat_200k-gpt-4-turbo-translated --split train_sft --split test_sft --columns prompt --src-lang English --tgt-lang Dutch --output-hub-name BramVanroy/ultrachat_200k_dutch --output-hub-revision 1-gpt-4-turbo-translated -j 8 --system-prompt .transl_sysprompt_en-nl

This will:

• Translate the train_sft and test_sft splits of HuggingFaceH4/ultrachat_200k from English to Dutch
• It will save temporary results to data/ultrachat_200k/ultrachat_200k-gpt-4-turbo-translated
• It will upload the final dataset to revision (branch) 1-gpt-4-turbo-translated in the BramVanroy/ultrachat_200k_dutch dataset
• It will use 8 processes to speed up the translation
• It will use the .transl_sysprompt_en-nl file that contains a system prompt as the system message

answer-hf

In the next step we want to use models or APIs to generate an answer to given columns. This script will do that for you. The only required input that is used is the given user-column as the user message, optionally a system-column, and the model answer to those will be saved into the response-column (defautls to response).

Example usage:

answer-hf --help

conversation-hf

This script allows you to build a conversation in a single model response. Importantly, the specified system_prompt is supposed to tell the model to create a multi-turn conversation and also give an example of such a conversation, with specified identifiers for the user and assistant in the generated conversation. These identifiers should also be given in this script (defaults to user: and assistant: ).

You can also specifiy personas with --personas which should be a JSON file containinga main key personas with persona names and their descriptions, which can then be passed to the system_prompt as long as it has a {persona} field in its text. The JSON file can optionally also have a weights key, which indicates how randomly weighted the different personas are chosen. If not given, all personas are equally likely. To repeat: when you provide a personas file, the persona descriptions will be randomly selected for each sample (optionally weighted) and plugged into the system_prompt that you provided as long as that text (file) contains the string {persona}.

Example usage:

answer-hf --help

interactive-lid

An interactive script to add language identification to specified columns in your dataset. The script handles messages (lists of dictionaries) by simply concatenating all content keys.

The script will add {colname}_lid and {colname}_lid_prob columns to your dataset.

Usage: simply run interactive-lid and follow instructions.

interactive-filter-dutch

An interactive script to filter out non-Dutch messages from your dataset. It does so based on the columns added with interactive-lid so that script should be used first.

In addition to language filtering, it also allows you to filter out messages with specific characteristics. Text matching occurs in a case-insensitive manner.

• messages with non-Roman characters are removed (every character must have "LATIN" in its unicode name; note that this solution is not flawless: https://stackoverflow.com/a/3308844/1150683) This is a very strict filter and will lead to the removal of data that you may have wanted to keep (e.g. messages that involve a translation task to non-Latin script languages)
• messages that are not identified as Dutch and that are longer than three white-space separated tokens are removed
• any text containing "spijt me", "spijt mij", "sorry", "mijn excuses", because those often indicate that the system could not successfully reply to a request
• any text containing "It seems like there was a typo", "assistant", because those often indicate that the system could not successfully reply to a request. Note that assistant is the English word (assistent is Dutch), so when assistant appears something is likely wrong
• any text indicating knowledge cut-offs:
  • kennisafsluiting in 2023
  • kennisstop in 2023
  • kennisafsnijdatum van 2023
  • cutoff in 2023
  • Tot mijn kennis die bijgewerkt is tot begin 2023
  • Voor zover mijn kennis reikt tot 2023
  • Vanaf mijn kennis tot begin 2023
  • As of my last update in 2023
• any text referencing other language models
  • ChatGPT
  • Chat GPT
  • GPT3
  • GPT 3
  • gpt-3
  • gpt-3.5-turbo
  • GPT4
  • GPT 4
  • gpt-4
  • gpt-4-turbo
  • OpenAI
  • ShareGPT
• any self-referencing text about being a language model. This often indicates that a model is not capable of a specific task, in case we drop those samples to instead focus on the tasks that it can do. The following strings are matched in a template for all occurrences of "als [een] {}", "ben [een] {}", "{} ben"
  • AI-assistent
  • AI-gebaseerde assistent
  • virtuele assistent
  • digitale assistent
  • tekst-assistent
  • AI tekstgebaseerde asssistent
  • tekstgebaseerde asssistent
  • assistent
  • taalmodel
  • AI-taalmodel
  • AI taalmodel

License

Licensed under GPLv3.