PLaMo 3 NICT 2604 31B Base
NOTE: This model has NOT been instruction-tuned for chat dialog or other downstream tasks. For instruction-tuned model, please use the PLaMo 3.0 Prime.
Model Description
PLaMo 3 NICT 2604 31B Base is a 31B model distilled from a 120B model, developed by Preferred Networks, Inc. in collaboration with National Institute of Information and Communications Technology, NICT.
PLaMo 3 NICT models adopt a hybrid architecture with Sliding Window Attention (SWA) and Traditional Attention layers.
PLaMo 3 NICT 2604 31B Base is released under PLaMo community license. Please check the following license and agree to this before downloading.
- (EN) https://plamo.preferredai.jp/info/plamo-community-license-en
- (JA) https://plamo.preferredai.jp/info/plamo-community-license-ja
For commercial users
Please check the PLaMo community license and contact us via the following form for commercial purpose.
Usage
Requirements
python>=3.10.0
numpy>=1.26.4
numba>=0.60.0
torch>=2.6.0,<=2.9.0
transformers>=5.0.0
Use vLLM
PLaMo3 is supported by vLLM.
import vllm
model = vllm.LLM("pfnet/plamo-3-nict-2604-31b-base", trust_remote_code=True)
response = model.generate(
prompts=["The future of artificial intelligence technology is "],
sampling_params=vllm.SamplingParams(
n=1,
max_tokens=32,
top_k=50,
top_p=0.95,
temperature=1.0,
),
)[0]
generated_text = response.outputs[0].text
print(generated_text)
Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("pfnet/plamo-3-nict-2604-31b-base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("pfnet/plamo-3-nict-2604-31b-base", trust_remote_code=True)
text = "これからの人工知能技術は"
input_ids = tokenizer(text, return_tensors="pt").input_ids
generated_tokens = model.generate(
inputs=input_ids,
max_new_tokens=32,
do_sample=True,
top_k=50,
top_p=0.95,
temperature=1.0,
)[0]
generated_text = tokenizer.decode(generated_tokens)
print(generated_text)
Model Details
- Model size: 31B
- Trained tokens: 1.5T
- Developed by:
- Preferred Networks, Inc.
- in collaboration with National Institute of Information and Communications Technology, NICT
- Model type: Causal decoder-only
- Language(s): English, Japanese
- License: PLaMo community license
Training Dataset
PLaMo 3 NICT 2604 31B was distilled from a 120B model which was trained with two phases, phase 1 with 2.625T tokens and phase 2 with 0.375T tokens. The percentage of datasets in each phase is shown in the following table.
| 2.625T (phase 1) | 0.375T (phase 2) | Tokens | |
|---|---|---|---|
| English | 45% | 35% | 1.3125T |
| Japanese | 30% | 40% | 0.9375T |
| Code | 15% | 15% | 0.4500T |
| Other | 10% | 10% | 0.3000T |
The training dataset also includes the following publicly available community resources:
- HuggingFaceFW/fineweb-2
- HuggingFaceTB/smollm-corpus
- bigcode/the-stack-v2-train-smol-ids
- nvidia/Nemotron-Pretraining-Dataset-sample
- EssentialAI/eai-taxonomy-stem-w-dclm
Tokenizer
PLaMo 3 NICT 2604 31B Base tokenizer is optimized by numba, which is a JIT compiler for numerical functions. The tokenizer is trained on a subset of the datasets for model pre-training.
Bias, Risks, and Limitations
PLaMo 3 NICT 2604 31B Base 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 3 NICT 2604 31B Base'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 3 NICT 2604 31B Base, developers should perform safety testing and tuning tailored to their specific applications of the model.