GRPO (main table, base+GRPO) — final checkpoint (SFT → GRPO). Base: Qwen/Qwen3-4B. Paper: Metacognitive Behavioral Tuning of Large Language Models for Multi-Hop Question Answering.

Method: GRPO: base Qwen3 trained directly with GRPO (no SFT stage).
Base model: Qwen/Qwen3-4B
Training: SFT (LR 1e-4, BS 128, HotpotQA) → GRPO
Benchmarks: HotpotQA (ID), MuSiQue / 2WikiMultiHopQA (OOD)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "metacognitive-behavioral-tuning/Qwen3-4B-GRPO"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")