README.md

---
language_creators:
- found
language:
- en
license: odc-by
source_datasets:
- c4
task_categories:
- text-generation
- fill-mask
task_ids:
- language-modeling
- masked-language-modeling
dataset_info:
  features:
  - name: text
    dtype: string
  - name: score
    dtype: float64
  splits:
  - name: train
    num_bytes: 373897649.51453334
    num_examples: 278115
  download_size: 242478448
  dataset_size: 373897649.51453334
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
size_categories:
- 100K<n<1M
---

# crumb/c4-benchfilter-nano

A derivation of the first 3M samples from the C4 dataset for a cheap and short continued pretraining for language models to optimize for benchmark scores without sacrificing generalization and generative modelling unrelated to chat or 'instruct' data. 

The estimated top 10% of highest estimated length normalized ngram (mean of tri, quad, and penta-gram) overlaps for each of the 
selected benchmark datasets (arc, truthful_qa, hellaswag, mmlu, humaneval) based 
on 1k samples, within the first 3M samples of C4. The top scoring sample 
datasets for each benchmark are then filtered again for top 30% scores and 
combined and exact-match de-duplicated. Then the top 3% scores and samples less than 200 characters long are removed
because they likely have exact large n-token matches by chance such as exact 
dates or times that aren't actually relevant to the data.\* 

\*Upon further examination, some of these samples are still present throughout the data, albeit at much lower frequency than before, you might benefit from using `dataset.filter(x['score'] > thresh)` for some threshold, but you risk losing high quality samples as well, this tradeoff should be well-examined before training.