Remek 🇵🇱
@remekai.bsky.social
2x Kaggle Grandmaster, Speakleash tech team (Polish LLM pre-training, finetuning and alignment). LLM model optimization (quantization, pruning), training pipeline optimization. Learning Triton and CUDA (I would like to be pro in this area).
Polish small language model called Bielik-1.5B. In Polish it outperforms all existing models. In English close to top performing. Will be available soon. We train it on Helios Supercomputer (AGH Cyfronet). 64x Grace Hooper 200 chips.
December 9, 2024 at 5:06 PM
Polish small language model called Bielik-1.5B. In Polish it outperforms all existing models. In English close to top performing. Will be available soon. We train it on Helios Supercomputer (AGH Cyfronet). 64x Grace Hooper 200 chips.
Bielik mini (alpha; Polish LLM) on my Raspberry Pi5! 1.5B (upscaled Llama-3.2-1B), replaced tokenizer (using auxiliary embedding method) and trained to prevent English forgetting. We will launch it soon (WIP).
November 28, 2024 at 5:04 PM
Bielik mini (alpha; Polish LLM) on my Raspberry Pi5! 1.5B (upscaled Llama-3.2-1B), replaced tokenizer (using auxiliary embedding method) and trained to prevent English forgetting. We will launch it soon (WIP).