metadata

license: mit
datasets:
  - sail/regmix-data
  - sail/regmix-data-sample
language:
  - en

Models Trained with Human Selection

This is a collection of the language models trained using Human selection, each with approximately 1B parameters, trained on different random mixtures of data. This project aims to validate the generalization capabilities of the RegMix approach (https://huggingface.co/papers/2407.01492) from small-scale (e.g., 1M parameters) to large-scale (e.g., 1B parameters) models.

Key Features

Model Size: 5 separate models trained with different seeds, each with ~1B parameters
Training Data: Human selection (from The Pile paper) data mixtures on the RegMix-Data dataset
Purpose: The Human selection is a strong baseline for our method RegMix

Dataset

The models were trained using the RegMix-Data dataset, which is split into different domains from The Pile dataset.

Training Hyperparameters

Hyperparameter	Value
Batch Size	1M tokens
Learning Rate	4e-4
Minimum Learning Rate	1e-5
Learning Rate Schedule	Cosine
Warmup Ratio	4%
Total Tokens	25B

How to Load a Model

You can load any model using the corresponding branch with the Hugging Face Transformers library:

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("sail/data-mixture-human-1b", revision="seed-1")
tokenizer = AutoTokenizer.from_pretrained("sail/data-mixture-human-1b", revision="seed-1")

Data Mixture

The specific data mixture used for training this 1B model is as follows, which can be also found in our code:

train:
  train_the_pile_arxiv: 0.1052
  train_the_pile_freelaw: 0.0386
  train_the_pile_nih_exporter: 0.0052
  train_the_pile_pubmed_central: 0.1071
  train_the_pile_wikipedia_en: 0.0919
  train_the_pile_dm_mathematics: 0.0198
  train_the_pile_github: 0.0427
  train_the_pile_philpapers: 0.0027
  train_the_pile_stackexchange: 0.0929
  train_the_pile_enron_emails: 0.0030
  train_the_pile_gutenberg_pg_19: 0.0199
  train_the_pile_pile_cc: 0.1121
  train_the_pile_ubuntu_irc: 0.0074
  train_the_pile_europarl: 0.0043
  train_the_pile_hackernews: 0.0075
  train_the_pile_pubmed_abstracts: 0.0845
  train_the_pile_uspto_backgrounds: 0.0420
valid:
  valid_the_pile_pile_cc: 1.0
model_name: tinyllama_1_1b

Domain weights will be normalized to make sure their sum is 1.0 for train sets in our code.

Model Variants

To access different model variants, simply change the revision parameter in the from_pretrained method to the desired seed (e.g., "seed-2", "seed-3"), and the maxium seed is 5.

Model Performance

We evaluated each model using lm-evaluation-harness. The performance metric for each task is the average of 0-shot to 5-shot accnorm (accuracy normalized, if available) or acc (accuracy) scores.

Seed	PIQA	LAMBADA	MultiRC	LogiQA	SocialIQA	Winogrande	RACE	OpenBookQA	COPA	HellaSwag	SciQ	ARC Easy	QQP	Average
1	65.00	29.83	54.28	25.47	33.61	53.06	28.98	28.17	66.67	37.43	80.13	49.40	52.42	46.50
2	65.03	26.69	53.24	25.31	33.69	52.52	29.42	28.76	63.00	37.68	82.58	51.36	58.46	46.75
3	65.57	28.47	54.18	25.68	34.24	52.31	30.12	28.00	65.80	37.90	82.48	49.34	56.53	46.97
4	65.45	26.88	51.42	24.92	34.16	50.50	29.93	28.92	62.40	37.70	80.66	49.27	58.06	46.17
5	66.67	29.56	51.58	26.94	33.22	51.78	29.03	28.56	65.00	37.69	81.78	50.38	52.60	46.52

Usage Notes

These models are primarily intended for research purposes.
Performance may vary depending on the specific task and domain.

Citation

If you use these models in your research, please cite the RegMix paper:

@article{liu2024regmix,
  title={RegMix: Data Mixture as Regression for Language Model Pre-training},
  author={Liu, Qian and Zheng, Xiaosen and Muennighoff, Niklas and Zeng, Guangtao and Dou, Longxu and Pang, Tianyu and Jiang, Jing and Lin, Min},
  journal={arXiv preprint arXiv:2407.01492},
  year={2024}
}

For more information about the RegMix methodology and its applications, please refer to the original paper.