recursive_self_improvement_dataset

# Dataset card ## This prompt and output data for recursive self improvement model: metatune-gpt20b-R1.1 - self generated prompt - 5 model checkpoint output - 5 generation responses ## Usage: 1) open source data for analyzing for improvement of the model 2) Benchmark improvement for model 3) Benchmark models ## Train **Do not train the dataset.** Only for benchmark! ## Risk: - Prompt safely with recursive self improvement model. Use safety gpt oss 20b for model safety analysis - Do not use this to prompt for nuclear, biological and chemical weapons. # Fine tune log: - [meta] Post-finetune eval_loss: 3.6989 improvement (loss delta): 0.8508 - [meta] Updated meta state -> lr=1.01e-04, epochs=1, patience=0 - [meta] Meta-loop complete. - Final meta state: {'learning_rate': 0.00010056785937499997, 'num_train_epochs': 1, 'total_fine_tune_steps_done': 97, 'patience_counter': 0}