curriculum learning - llama3-8b-instruct finetuned on easy/hard stages of clem games, with hard stage being the (game,exp,task_id) tuples the base model failed at.
this is the model trained on both the easy stage (successful task_ids of base model) and the hard stage (failed task_ids)
batch_size: 4
learning_rate: 0.0002
max_steps: 500
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
