MLDataScientist's picture
Update README.md
4b8802f verified
|
raw
history blame
No virus
1.72 kB
metadata
license: apache-2.0
dataset_info:
  features:
    - name: text
      dtype: string
  splits:
    - name: validation
      num_bytes: 642375
      num_examples: 535
    - name: train
      num_bytes: 15585375
      num_examples: 12703
  download_size: 7315916
  dataset_size: 16227750
configs:
  - config_name: default
    data_files:
      - split: validation
        path: data/validation-*
      - split: train
        path: data/train-*

Open Assistant Conversations Dataset Release 2 (OASST2) in Uzbek language

This dataset is an Uzbek translated version of OASST2 dataset in a thread format with Llama3 chat template.

Refer to this translated version if you need the original tree format. Otherwise, use this thread format for fine-tuning Llama3 models.


The Uzbek translation was completed in 45 hours using a single T4 GPU and nllb-200-3.3B model.

Based on nllb metrics, you might want to only filter out records that were not originally in English or Russian since English-Uzbek and Russian-Uzbek have acceptable metrics and translation quality is noticeable better for those pairs based on my short reviews.

I am sharing the entire Uzbek translated dataset for future research.

The following repo and command was used to do the Uzbek translation.

Repo: https://github.com/UnderstandLingBV/LLaMa2lang

Command used:

!python3 translate.py nllb --model_size 3.3B uzn_Latn output_uzbek --quant8 --base_dataset OpenAssistant/oasst2 --max_length 512 --checkpoint_n 400 --batch_size 40

I will fine-tune LLAMA3 8B Uzbek chat model and release in HF soon.