Model: gczhu/Qwen3-4B-Warmup Source: Original Platform
license, base_model, library_name, pipeline_tag, tags
| license | base_model | library_name | pipeline_tag | tags | |||||
|---|---|---|---|---|---|---|---|---|---|
| apache-2.0 |
|
transformers | text-generation |
|
PivotTrace — Qwen3-4B Formatting Warmup
This repository provides the formatting-warmup checkpoint of Qwen3-4B-Base used in our NeurIPS 2026 paper:
Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots.
This checkpoint serves as the starting point for PivotTrace data selection and subsequent RLVR. We release it so that users can skip formatting warmup and proceed directly to reproducing our method.
Model Details
- Base model: Qwen/Qwen3-4B-Base
- Training stage: Formatting warmup
- Intended use: PivotTrace data selection and subsequent RLVR training
- Code and training recipe: gczhu/PivotTrace
Getting Started
Follow the installation instructions in the code repository, then download the contents of this model repository into:
PivotTrace/models/Qwen3-4B-Warmup/
Place the configuration, tokenizer files, and model weights directly inside this directory. The selection and training scripts use this checkpoint path by default.
From the root of the code repository, run:
bash selection/01_generate_response.sh # Generate reasoning trajectories
bash selection/02_score_data.sh # Trace pivots and score questions
bash selection/03_active_select.sh # Calibrate thresholds and select data
bash train_scripts/run_semi_train.sh # Train on labeled and unlabeled data
The processed datasets and full experimental settings are provided in the code repository.
Citation
If you find this checkpoint or our code useful, please consider citing:
@article{zhu2026smart,
title = {Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots},
author = {Zhu, Guangcheng and Yang, Shenzhi and Wang, Haobo and Zheng, Xing and Ma, Yingfan and Feng, Xuening and Chen, Zhongqi and Song, Bowen and Wang, Weiqiang and Chen, Gang},
journal = {arXiv preprint arXiv:2606.04503},
year = {2026},
doi = {10.48550/arXiv.2606.04503},
url = {https://arxiv.org/abs/2606.04503}
}