it was used to fine-tune TinyLlama/TinyLlama-1.1B-Chat-v1.0 for tweet tone classification problem. Default model achieved 0.20 f1-score, while fine-tuned version achieved 0.55 f1-score in less than 4 minutes of fine-tuning on a single A100

LoRA was used with r=16 and alpha=32 to fine-tune "q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "down_proj".

BATCH_SIZE = 128 LEARNING_RATE = 1e-3 NUM_EPOCHS = 1

F1 score is 0.54 on a test set