reasoning-degeneration-dev/

multiplex-countdown-configs-v1

Training configurations for multiplex-thinking-countdown experiment. Two conditions: multiplex (w=3, soft thinking) vs baseline CoT (vanilla GRPO). Both train DeepSeek-R1-Distill-Qwen-7B on Countdown d4-d5.

Rows: 2 Columns: 26

	Columns




	

Column Type Description

condition Value('string') Experimental condition: multiplex or baseline_cot

experiment_name Value('string') WandB experiment name

loss_mode… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/multiplex-countdown-configs-v1.