---
license: mit
base_model: meta-llama/Llama-3.2-3B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags: [chain-of-thought, gsm8k, cot, sft]
---

# gsm-cot-llama3b-nl

Natural-language chain-of-thought SFT checkpoint, fine-tuned from `meta-llama/Llama-3.2-3B-Instruct`
on GSM8k-Aug with **natural-language** reasoning steps. From the paper
[Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers](https://arxiv.org/abs/2606.31779).

This is the **Stage-1 CoT initialization** for the natural-language LOTUS model
([`yingfanbot/gsm-lotus-llama3b-nl`](https://huggingface.co/yingfanbot/gsm-lotus-llama3b-nl)).

It is a standard causal LM — load with `from_pretrained` and generate the CoT directly (no wrapper needed).

## Citation
```bibtex
@article{fan2026bridging,
  title={Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers},
  author={Fan, Ying and Svete, Anej and Lee, Kangwook},
  journal={arXiv preprint arXiv:2606.31779},
  year={2026}
}
```