metacognitive-behavioral-tuning/Qwen3-4B-MBT-R

Qwen3-4B · MBT-R

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

• —Method: MBT-R (Refinement): the student's own reasoning traces are rewritten into the 5-phase structure for SFT, then GRPO.
• —Base model: Qwen/Qwen3-4B
• —Training: SFT (LR 1e-4, BS 128, HotpotQA) → GRPO
• —Benchmarks: HotpotQA (ID), MuSiQue / 2WikiMultiHopQA (OOD)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

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