apptainer shell --nv ./kulyk.sif

Apptainer> /opt/entrypoints/kulyk --verbose --n-len 1024 --model-path-ue /project/models/kulyk-uk-en.gguf --model-path-eu /project/models/kulyk-en-uk.gguf

Run it as a webservice:

apptainer instance start --nv ./kulyk.sif kulyk-ws

# go to http://localhost:3000

Facts:

Fine-tuned with 40M samples (filtered by quality metric) from ~53.5M for 1.4 epochs
354M params
Requires 1 GB of RAM to run with bf16
BLEU on FLORES-200: 27.24
Tokens per second: 229.93 (bs=1), 1664.40 (bs=10), 8392.48 (bs=64)
License: lfm1.0

Info:

Model name is inherited from name of Sergiy Kulyk who was chargé d'affaires of Ukraine in the United States

Training Info:

Learning Rate: 3e-5
Learning Rate scheduler type: cosine
Warmup Ratio: 0.05
Max length: 2048
Batch Size: 10
packed=True
Sentences <= 1000 chars
Gradient accumulation steps: 4
Used Flash Attention 2
Time for epoch: 32 hours
2 cards of NVIDIA RTX 3090 Ti (24G)
accelerate with DeepSpeed, offloading into CPU
Memory usage: 22.212GB-22.458GB
torch 2.7.1

Acknowledgements:

Serhiy Stetskovych for providing compute to train this model
lang-uk members for their compilation of different MT datasets

Yehor/kulyk-en-uk

Yehor/
kulyk-en-uk