set unique UNK token

[Ssukriti]

Description of the change

Granite models have UNK = EOS , this is resulting in poor quality when tuning for some datasets. When it is set to unique, the quality improves.

Related issue number

https://github.ibm.com/ai-foundation/watson-fm-stack-tracker/issues/1435

How to verify the PR

Tested with tone dataset
/home/tuning/.local/bin/accelerate launch --num_processes=2 --config_file /app/accelerate_fsdp_defaults.yaml -m tuning.sft_trainer --model_name_or_path $MODEL_PATH --training_data_path $TRAIN_DATA_PATH --torch_dtype bfloat16 --output_dir $OUTPUT_PATH --num_train_epochs 5 --per_device_train_batch_size 4 --gradient_accumulation_steps 4 --learning_rate 1e-5 --response_template "\n### Response:" --dataset_text_field "output"

export MODEL_PATH="ibm-granite/granite-3.0-8b-base"
export TRAIN_DATA_PATH="/testing/tuning/input/cc_tone_sft_format_1000_train.json"

at inference we get repeated output without the change , and proper output after change

Was the PR tested

• I have added >=1 unit test(s) for every new method I have added.
• I have ensured all unit tests pass

[github-actions]

Thanks for making a pull request! 😃
One of the maintainers will review and advise on the next steps.