outputs

This model is a fine-tuned version of microsoft/phi-2 using trl on ultrafeedback dataset.

What's new

A test for ORPO: Monolithic Preference Optimization without Reference Model method using trl library.

How to reproduce

accelerate launch --config_file=/path/to/trl/examples/accelerate_configs/deepspeed_zero2.yaml \
    --num_processes 8 \
    /path/to/trl/scripts/orpo.py \
    --model_name_or_path="microsoft/phi-2" \
    --per_device_train_batch_size 1 \
    --max_steps 8000 \
    --learning_rate 8e-5 \
    --gradient_accumulation_steps 1 \
    --logging_steps 20 \
    --eval_steps 2000 \
    --output_dir="orpo-lora-phi2" \
    --optim rmsprop \
    --warmup_steps 150 \
    --bf16 \
    --logging_first_step \
    --no_remove_unused_columns \
    --use_peft \
    --lora_r=16 \
    --lora_alpha=16 \
    --dataset HuggingFaceH4/ultrafeedback_binarized