Difference between revisions of "Lumi/pilot"
(→Data: Norwegian) |
(→Model Architectures) |
||
| Line 14: | Line 14: | ||
= Model Architectures = | = Model Architectures = | ||
| + | * [https://arxiv.org/pdf/1910.10683.pdf T5] | ||
| + | * Ablations with BERT | ||
| + | * [[Eosc/pretraining#ELECTRA|ELECTRA]] | ||
* [[Eosc/pretraining#BERT|BERT]] (separate Bokmål and Nynorsk models) | * [[Eosc/pretraining#BERT|BERT]] (separate Bokmål and Nynorsk models) | ||
* [[Eosc/pretraining#RoBERTa|RoBERTa]] | * [[Eosc/pretraining#RoBERTa|RoBERTa]] | ||
| − | |||
* [https://openai.com/blog/gpt-2-1-5b-release/ GPT] | * [https://openai.com/blog/gpt-2-1-5b-release/ GPT] | ||
| − | |||
* Large language models with linguistically motivated inductive biases (linked to the dScience PhD position); one example is Google's [https://www.aclweb.org/anthology/2020.emnlp-main.19/ ETC]. | * Large language models with linguistically motivated inductive biases (linked to the dScience PhD position); one example is Google's [https://www.aclweb.org/anthology/2020.emnlp-main.19/ ETC]. | ||
Revision as of 12:17, 1 June 2022
LUMI-G Pilot
In late 2021, the shared LUMI supercomputer will (likely) open for trial usage of its vast gpu partition. NLPL partners in Finland (Turku and Helsinki) and Norway (Oslo) are coordinating their efforts towards the creation of very large-scale (neural) language models for multiple Nordic languages. This work is part of the Nordic Language Modeling (NorLM) initiative.
Model Architectures
- T5
- Ablations with BERT
- ELECTRA
- BERT (separate Bokmål and Nynorsk models)
- RoBERTa
- GPT
- Large language models with linguistically motivated inductive biases (linked to the dScience PhD position); one example is Google's ETC.
Software Support
See the links above for particular model's requirements.
In general, we rely on Python (>=3.7) and its SciPy stack.
We definitely will require fully functional GPU-enabled installations of PyTorch (1.8.1) and TensorFlow (preferably, both 1.15.5 and 2.4.1).
Multi-GPU and multi-node training must be possible. In the NVIDIA world, NCCL and Horovod are used for this. In the AMD world? No idea.
Data: Norwegian
- Collaboration with the National Library (Colossal Norwegian Corpus)?
- Extracting the Norwegian part from the C4 dataset?
- Additional news collections (Lilja?)
