README.md

metadata

license: cc
tags:
  - custom
  - subset

This dataset is a subset of allenai/Dolci-Instruct-SFT.

• Source revision: 51b25163854ed9da85df78c7e9267ecbaf9c36a0
• Tokenizer revisions: {"Qwen/Qwen3-8B-Base": "49e3418fbbbca6ecbdf9608b4d22e5a407081db4", "allenai/OLMo-2-0425-1B": "a1847dff35000b4271fa70afc5db10fd29fedbdf", "allenai/Olmo-3-1025-7B": "a81bae42db3975be1671e27b9c9a56da1a9f980f", "meta-llama/Llama-3.2-1B": "4e20de362430cd3b72f300e6b0f18e50e7166e08"}

Filtering

• Sample first, validate candidates, and replace rejected rows.
• Rejection counts cover examined candidates only.
• Conversation validation: enabled
• Maximum tokens: 2048 (tokenizers: Qwen/Qwen3-8B-Base, allenai/Olmo-3-1025-7B, allenai/OLMo-2-0425-1B, meta-llama/Llama-3.2-1B, template: ./chat_template.jinja)
• train: 49,066 candidates examined of 2,152,112 original; 6,024 rejected, 43,042 accepted
  • invalid_role: 4,457
  • invalid_content_type: 4,688
  • empty_content: 26
  • too_many_tokens:Qwen/Qwen3-8B-Base: 1,257
  • too_many_tokens:allenai/Olmo-3-1025-7B: 1,181
  • too_many_tokens:allenai/OLMo-2-0425-1B: 1,181
  • too_many_tokens:meta-llama/Llama-3.2-1B: 1,138

Dataset Splits

• Train: 21,521 samples (1.00% of original 'train' split; 49,066 candidates examined)
• Test: 21,521 samples (1.00% of original 'train' split; 49,066 candidates examined)
• Seed: 42

Generation Command

python take_split.py --source_dataset allenai/Dolci-Instruct-SFT --target_dataset loris3/Dolci-Instruct-SFT-sample --train 0.01 --test 0.01 --seed 42 --num_proc 8 --source_revision 51b25163854ed9da85df78c7e9267ecbaf9c36a0 --tokenizer Qwen/Qwen3-8B-Base --tokenizer allenai/Olmo-3-1025-7B --tokenizer allenai/OLMo-2-0425-1B --tokenizer meta-llama/Llama-3.2-1B --validate --chat_template ./chat_template.jinja --max_tokens 2048 --private

License

See original dataset.