smoltldr / README.md
mlabonne's picture
Update README.md
b6d3836 verified
metadata
language:
  - en
dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: completion
      dtype: string
  splits:
    - name: train
      num_bytes: 3111855.6398965064
      num_examples: 2000
    - name: validation
      num_bytes: 311390.6933457422
      num_examples: 200
    - name: test
      num_bytes: 311850.6638180986
      num_examples: 200
  download_size: 1739330
  dataset_size: 3735096.9970603473
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

This dataset was designed for the fine-tune of HuggingFaceTB/SmolLM2-135M-Instruct using GRPO. It is designed to summarize Reddit posts.

You can reproduce this training using this colab notebook. It takes about 40 minutes to train the model.