---
model-index:
- name: Kulyk-UK-EN
  results:
  - task:
      type: text-generation
    dataset:
      type: facebook/flores
      name: FLORES
      split: devtest
    metrics:
    - type: bleu
      value: 36.27
      name: BLEU
library_name: transformers
license: other
license_name: lfm1.0
license_link: LICENSE
language:
- en
- uk
pipeline_tag: text-generation
tags:
- liquid
- lfm2
- edge
datasets:
- lang-uk/FiftyFiveShades
base_model:
- LiquidAI/LFM2-350M
---

A lightweight model to do machine translation from Ukrainian to English based on recently published LFM2 model. Use [demo](https://huggingface.co/spaces/Yehor/uk-en-translator) to test it.

Also, there's another model: [kulyk-en-uk](https://huggingface.co/Yehor/kulyk-en-uk)

**Run with Docker (CPU)**:

```
docker run -p 3000:3000 --rm ghcr.io/egorsmkv/kulyk-rust:latest
```

**Run using Apptainer (CUDA)**:

1. Run it using shell:

```
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
```

2. 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: 36.27
- 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](https://en.wikipedia.org/wiki/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
- Memory usage: 22.212GB-22.458GB
- torch 2.7.1

**Acknowledgements**:

- [Dmytro Chaplynskyi](https://huggingface.co/dchaplinsky) for providing compute to train this model
- [lang-uk](https://huggingface.co/lang-uk) members for their compilation of different MT datasets