OLMo
In this guide, we provide an overview of the Open Language Mode (OLMo), including prompts and usage examples. The guide also includes tips, applications, limitations, papers, and additional reading materials related to OLMo.
Introduction to OLMo
The Allen Institute of AI has released (opens in a new tab) a new open language model and framework called OLMo. This effort is meant to provide full access to data, training code, models, evaluation code so as to accelerate the study of language models collectively.
Their first release includes four variants at the 7B parameter scale and one model at the 1B scale, all trained on at least 2T tokens. This marks the first of many releases which also includes an upcoming 65B OLMo model.
The releases includes:
- full training data, including the code (opens in a new tab) that produces the data
- full models weights, training code (opens in a new tab), logs, metrics, and inference code
- several checkpoints per model
- evaluation code (opens in a new tab)
- fine-tuning code
All the code, weights, and intermediate checkpoints are released under the Apache 2.0 License (opens in a new tab).
OLMo-7B
Both the OLMo-7B and OLMo-1B models adopt a decoder-only transformer architecture. It follows improvements from other models like PaLM and Llama:
- no biases
- a non-parametric layer norm
- SwiGLU activation function
- Rotary positional embeddings (RoPE)
- a vocabulary of 50,280
Dolma Dataset
This release also includes the release a pre-training dataset called Dolma (opens in a new tab) -- a diverse, multi-source corpus of 3 trillion token across 5B documents acquired from 7 different data sources. The creation of Dolma involves steps like language filtering, quality filtering, content filtering, deduplication, multi-source mixing, and tokenization.
The training dataset includes a 2T-token sample from Dolma. The tokens are concatenated together after appending a special EOS token to the end of each document. The training instances include groups of consecutive chunks of 2048 tokens, which are also shuffled.
More training details and hardware specifications to train the models can be found in the paper.
Results
The models are evaluated on downstream tasks using the Catwalk (opens in a new tab). The OLMo models are compared to other several publicly available models like Falcon and Llama 2. Specifically, the model is evaluated on a set of tasks that aim to measure the model's commonsense reasoning abilities. The downstream evaluation suite includes datasets like piqa and hellaswag. The authors perform zero-shot evaluation using rank classification (i.e., completions are ranked by likelihood) and accuracy is reported. OLMo-7B outperforms all other models on 2 end-tasks and remains top-3 on 8/9 end-tasks. See a summary of the results in the chart below.
Prompting Guide for OLMo
Coming soon...
Figures source: OLMo: Accelerating the Science of Language Models (opens in a new tab)