Using custom monolingual data instead of OSCAR dataset

[learnercat]

Thank you for your great work. I am preparing to start an experiment with the ALMA for low-resource languages. I want to use and fine-tune the monolingual model with my local datasets in stage 1 instead of OSCAR dataset. Should I follow the OSCAR JSON file format as below? Please advise.

-km train_raw.json valid_raw.json test_raw.json

-th -ja

[{‘id': 0, 'text': “TEXT”, “lang”: “LANG”}, {‘id': 0, 'text': “TEXT”, “lang”: “LANG”}, {‘id': 0, 'text': “TEXT”, “lang”: “LANG”}, …. ]

[fe1ixxu]

Thanks for the interest!

The easiest way is to pass the flag --mono_data_path $YOUR_MONO_DATA to the training command, e.g.,

accelerate launch --config_file configs/deepspeed_train_config.yaml \
     run_llmmt.py \
    --model_name_or_path meta-llama/Llama-2-7b-hf \
    --mono_data_path $YOUR_MONO_DATA \
    --max_steps 600000 \
    --do_train \
    --low_cpu_mem_usage \
    --fp16 \
    --learning_rate 2e-5 \
    --weight_decay 0.01 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --warmup_ratio 0.01 \
    --ignore_pad_token_for_loss \
    --ignore_prompt_token_for_loss \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --save_strategy steps \
    --save_steps 2000 \
    --save_total_limit 1 \
    --logging_strategy steps \
    --logging_steps 1 \
    --output_dir ${OUTPUT_DIR} \
    --max_new_tokens 256 \
    --max_source_length 256 \
    --seed 42 \
    --overwrite_output_dir \
    --report_to none

The format of your mono data should follow the translation parallell data, but leave the English sentence empty, e.g.,

{
  "translation":
  {
      "de": "mono sentence",
      "en": "",
}
}

[learnercat]

Thank you very much for the complete guide for Monolingual training.