Pilot counterfactual-importance rollout trees (remove chunk i, resample, measure downstream impact) on MGSM/MMATH/MMMLU across API models, Target-CoT conditions.
models
meta-llama/Llama-3.2-3B-Instruct-Turbo, Qwen/Qwen3.5-9B, Qwen/QwQ-32B, Qwen/Qwen2.5-1.5B-Instruct, DeepSeek-R1-Distill-Qwen-14B, deepseek-r1-distill-qwen-7b… See the full description on the dataset page:
https://huggingface.co/datasets/multicot/2026-04-30-rollout-importance-pilot-rollouts.