zisisbatzos/GRPO_3SFTs_synthetic_emobench_llama3.2-3B

training_args = GRPOConfig( output_dir=output_dir, per_device_train_batch_size=4, gradient_accumulation_steps=4, num_generations=8, max_prompt_length=256, max_completion_length=512, learning_rate=1e-5, num_train_epochs=10, logging_steps=1, save_strategy="steps", save_steps=100, use_vllm=False, save_safetensors=False, max_grad_norm=1.0, ddp_find_unused_parameters=False, )

baseline_model = 'zisisbatzos/3SFTs_synthetic_emobench_llama3.2-3B'

Metrics: emotion_accuracy = 0.7563025210084033 correct_format = 1.0