Interplay-LM-Reasoning/context_pretrain_2

Interplay-LM Context Pretrain 2

This repository contains the context B pretraining checkpoints and the corresponding final RL checkpoints. In this setting, the teacher component uses only op2 during pretraining.

Only inference-relevant Hugging Face files are included.

Within each setting:

• base/ stores the final op2-only pretraining checkpoint.
• rl/ stores the final RL checkpoints for each experiment variant.

Included settings

• 0.9zoo_op2-20+0.1teacher_op2
• 0.99zoo_op2-20+0.01teacher_op2
• 0.999zoo_op2-20+0.001teacher_op2

Load

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Interplay-LM-Reasoning/context_pretrain_2"
subdir = "0.99zoo_op2-20+0.01teacher_op2/rl/contextzoo_0.99zoo_0.01teacher_process_strict"

tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subdir)
model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder=subdir)

Citation

@misc{zhang2025interplaypretrainingmidtrainingrl,
      title={On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models},
      author={Charlie Zhang and Graham Neubig and Xiang Yue},
      year={2025},
      eprint={2512.07783},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.07783},
}