PLaMo 2 1B is a 1B model pre-trained on English and Japanese datasets, developed by Preferred Elements, Inc.
PLaMo 2 models adapt the hybrid architecture like Samba rather than the Transformer architecture. Samba integrates Mamba, a selective State Space Model (SSM), with sliding window attention, combining their strengths for improved efficiency and performance. The major differences between Samba and PLaMo 2 are 1) adding normalization layers to improve training stability, and 2) using Mamba2 kernel for computational efficiency.
PLaMo 2 1B is released under Apache License version 2.0.
NOTE: This model has NOT been instruction-tuned for chat dialog or other downstream tasks.
We trained PLaMo 2 1B in two phases, phase 1 with 3.5T tokens and phase 2 with 0.5T tokens.
The percentage of datasets in each phase is shown in the following table.
3.5T (phase 1)
0.5T (phase 2)
Tokens
English
45 %
35 %
1.75 T
Japanese
30 %
40 %
1.25 T
Coding
15 %
15 %
0.6 T
Other
10 %
10 %
0.4 T
Tokenizer
PLaMo 2 1B tokenizer is optimized by numba, which is JIT compiler for numerical functions.
The tokenizer is trained on a subset of the datasets for model pre-training.
PLaMo 2 1B is a new technology that carries risks with use. Testing conducted to date has been in English and Japanese, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, PLaMo 2 1B’s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of PLaMo 2 1B, developers should perform safety testing and tuning tailored to their specific applications of the model.
Acknowledgement
This model is trained under the project, “Research and Development Project of the Enhanced Infrastructures for Post 5G Information and Communication System” (JPNP 20017), subsidized by the New Energy and Industrial Technology Development Organization (NEDO).
