SlimPajama-DC is a set of 1.3B parameter language models, distinctively trained on the different combinations of 330B subsets of SlimPajama dataset.

Details of Dataset Combinations for Different Models

Despite being trained on a smaller amount of 330B tokens compared to TinyLlama and Olmo's 3 trillion, SlimPajama-DC surpasses TinyLlama and Olmo in some challenging English tasks.

Our tests comprise: (1) AI2 Reasoning Challenge (25-shot); (2) HellaSwag (10-shot); (3) MMLU (5-shot); (4) TruthfulQA (0-shot)

‡ represents the RefinedWeb CC.

Performance on More Benchmarks

ARC easy and ARC challenge are evaluated using 25-shot. All other evaluation benchmarks are tested on 0-shot. * represents the results are averaged across multiple sub-items inside each benchmark dataset.

Dataset

Our full dataset is available at SlimPajama-627B-DC.

Model Usage

To load a specific checkpoint, use the revision argument as shown below, for example, SlimPajama-DC-6. All the revisions can be seen from the branch dropdown in the "Files and versions" tab. If no revision argument is provided, it will load the default checkpoint SlimPajama-DC-6.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(
    "MBZUAI-LLM/SlimPajama-DC",
    revision="SlimPajama-DC-6",
    trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    "MBZUAI-LLM/SlimPajama-DC",
    revision="SlimPajama-DC-6",
    trust_remote_code=True
)

prompt = 'int add(int x, int y) {'

input_ids = tokenizer(prompt, return_tensors="pt").input_ids
gen_tokens = model.generate(input_ids, do_sample=True, max_length=400)

print("-"*20 + "Output for model"  + 20 * '-')
print(tokenizer.batch_decode(gen_tokens)[0])

Citation

BibTeX:

bibtex

@article{shen2023slimpajama,
  title={Slimpajama-dc: Understanding data combinations for llm training},
  author={Zhiqiang Shen, Tianhua Tao, Liqun Ma, Willie Neiswanger, Zhengzhong Liu, Hongyi Wang, Bowen Tan, Joel Hestness, Natalia Vassilieva, Daria Soboleva, Eric Xing},
  journal={arXiv preprint arXiv:2309.10818},
  year={2023}
}