🦙🗽 Small research project - how much it would cost to create Alpaca-like dataset, with 50k+ demonstrations, using slightly different approach. All data byproducts are CC0/MIT-licensed.

🔥 The project also contains 100k+ MIT-licensed demonstrations from Anthropics HH-RLHF repo - converted into "Alpaca compatible format".

👉 Follow me on Twitter for news and updates.

🚫 Remember that releasing a model based on data you generated via model API might violate the Terms of Service of the model API provider.

BTW: This repo shows how easy it is to fine-tune (PEFT=LORA) Flan-T5-* model with Alpaca-like dataset.

1. Clone the repo: git clone https://github.com/mobarski/alpaca-libre && cd alpaca-libre
2. Install required python modules: pip install -r requirements.txt
3. View / edit generate.py
4. Set API_KEY: export OPENAI_KEY=...
5. Run the script: python3 generate.py

• data/seed_tasks.jsonl - is from the Self-Instruct paper
• data/alpaca_libre_prompt_v1.txt - is from the Alpaca paper (with slight modfification)

Files in the data/output directory are in the same format as original Alpaca dataset.

Files in the data/output/work directory are in the .jsonl format and:

• contain one task (JSON object) per line,

• contain also tasks that failed quality checks (status!='ok')

  • these tasks might be marked as 'ok' after manual inspection

• each task object has the following items:

  • status - anything other than 'ok' is bad

  • instruction - instruction part of the prompt

  • input - input part of the prompt

  • output - expected output

  • other - dictionary for other information (similarity, etc)

GitHub repos:

• https://github.com/tatsu-lab/stanford_alpaca
• https://github.com/yizhongw/self-instruct
• https://github.com/orhonovich/unnatural-instructions
• https://github.com/gururise/AlpacaDataCleaned
• https://github.com/anthropics/hh-rlhf
• https://github.com/Reason-Wang/flan-alpaca-lora

Papers:

• https://crfm.stanford.edu/2023/03/13/alpaca.html
• https://arxiv.org/abs/2212.10560
• https://arxiv.org/abs/2212.09689
• https://arxiv.org/abs/2204.05862

• 0.4.2
  • MIT-licensed demonstrations from Anthropics HH-RLHF repo
    • 104k human preferred responses from the train datasets:
      • 41k harmless
      • 42k helpful
      • 21k helpful-online
• 0.4.1
  • v4 dataset converted into the same format as original Alpaca
  • jsonl dataset moved into work dir
• 0.4
  • grouping turns into rounds
  • basic input quality check
  • better <noinput> handling
  • <nooutput> handling
  • retry with backoff on API error
  • progressbars
  • fixed: typos in Alpaca prompt
  • fixed: whitespace handling after task number
• 0.3
  • parallel main loop
  • better cli output
  • output format change (everythig not essential is placed in the "other" object)
  • basic output quality check
  • fixed: multiline input/output handling
  • fixed: no initial space / empty section handling
  • fixed: <noinput>