ibm-granite/granite-7b-base

Model Name: Granite-7b-base

License: Apache-2.0

Languages: Primarily English

Architecture: The model architecture is a replica of Meta’s Llama2-7B base variant with MHA, trained with 1M batch size on 2T tokens.

Context Length: 4k tokens

Tokenizer: Llama2

Model Developers: IBM Research

Representing IBM’s commitment to open source innovation IBM has released granite-7b-base, a base pre-trained LLM from IBM’s Granite model series, under an apache-2.0 license for community and commercial use. Granite-7b-base was pre-trained from scratch on IBM-curated data as an open reference implementation of Meta’s Llama-2-7B. In a commitment to data transparency and fostering open innovation, the data sources, sampling proportions, and URLs for access are provided below.

For more information about training this model, please check out the blog: https://pytorch.org/blog/maximizing-training/

Pre-Training Data

The model was trained on 2T tokens, with sampling proportions designed to match the sampling distributions released in the Llama1 paper as closely as possible.

Dataset	Description	Sampling Proportion	URL
Common Crawl	Open repository of web crawl data with snapshots ranging from 2021 to 2023.	77%	https://data.commoncrawl.org/
Github_Clean	Code data from CodeParrot covering a variety of coding languages.	5.50%	https://huggingface.co/datasets/codeparrot/github-code-clean
Wikipedia and Wikimedia	Eight Wikimedia projects (enwiki, enwikibooks, enwikinews, enwikiquote, enwikisource, enwikiversity, enwikivoyage, enwiktionary). containing extracted plain text from pages and articles.	2%	https://dumps.wikimedia.org
USPTO	US patents granted from 1975 to May 2023, excluding design patents.	5%	https://bulkdata.uspto.gov/
PubMed Central	Biomedical and life sciences papers.	1.75%	https://ftp.ncbi.nlm.nih.gov/pub/pmc/oa_package/
arXiv	Over 1.8 million scientific paper pre-prints posted to arXiv.	2.50%	https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T
StackExchange	Anonymized set of all user-contributed content on the Stack Exchange network, a popular collection of websites centered around user-contributed questions and answers.	1%	https://archive.org/details/stackexchange_20221206
PG19	A repository of free e-books with focus on older works for which U.S. copyright has expired.	0.25%	https://github.com/google-deepmind/pg19
Webhose	Unstructured web content converted into machine-readable data feeds purchased by IBM.	5%	N/A

Evaluation Results

LM-eval Harness Scores

Evaluation metric	Llama2-7B (baseline)	Granite-7b-base
MMLU (zero shot)	0.41	0.43
MMLU (5-shot weighted avg)	0.47	0.50
Arc challenge	0.46	0.44
Arc easy	0.74	0.71
Boolq	0.78	0.76
Copa	0.87	0.83
Hellaswag	0.76	0.74
Openbookqa	0.44	0.42
Piqa	0.79	0.79
Sciq	0.91	0.91
Winogrande	0.69	0.67
Truthfulqa	0.39	0.39
GSM8k (8-shot)	0.13	0.11

Bias, Risks, and Limitations

Granite-7b-base is a base model and has not undergone any safety alignment, there it may produce problematic outputs. In the absence of adequate safeguards and RLHF, there exists a risk of malicious utilization of these models for generating disinformation or harmful content. Caution is urged against complete reliance on a specific language model for crucial decisions or impactful information, as preventing these models from fabricating content is not straightforward. Additionally, it remains uncertain whether smaller models might exhibit increased susceptibility to hallucination in ungrounded generation scenarios due to their reduced sizes and memorization capacities. This aspect is currently an active area of research, and we anticipate more rigorous exploration, comprehension, and mitigations in this domain.