guanning/CoT_Premature_Confidence

CoT Premature Confidence — Released Checkpoints

Checkpoints for the paper Understanding and Mitigating Premature Confidence for Better LLM Reasoning. Source repo: guanning03/CoT_Premature_Confidence.

Layout

countdown/
  4_30_100/oc_0.0/   GRPO on countdown-4-30-100, overconf coefficient 0.0
  4_30_100/oc_1.0/   GRPO on countdown-4-30-100, overconf coefficient 1.0
  4_10_50/oc_0.0/    GRPO on countdown-4-10-50,  overconf coefficient 0.0
  4_10_50/oc_1.0/    GRPO on countdown-4-10-50,  overconf coefficient 1.0
math/
  oc_0.0/            GRPO on dapo_hard, overconf coefficient 0.0
  oc_1.0/            GRPO on dapo_hard, overconf coefficient 1.0

oc_X.Y denotes the value of probe.overconf_coeff used during training.

• Countdown checkpoints are Qwen2.5-3B fine-tunes in HuggingFace safetensors format and can be loaded directly with AutoModelForCausalLM.from_pretrained(...).
• Math checkpoints are Qwen2.5-Math-7B FSDP shards (model_world_size_N_rank_*.pt). Use the verl checkpoint merger (scripts/model_merger.py in the source repo) to consolidate them into a single HuggingFace-loadable model.