README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - on-policy-distillation
  - cross-tokenizer
  - knowledge-distillation
  - mathematics
  - code
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files: mix-20k.jsonl

BPM Training Prompts (mix-20k)

arXiv:2607.22334 · Project page

Prompt corpus behind every reported result in BPM (Byte-Prefix Marginalization), a cross-tokenizer on-policy distillation method. Prompt-only.

Domain	Rows	Source	Upstream license
Mathematics	10,000	dapo-17k	Apache-2.0
Code	10,000	taco:*	Apache-2.0

Fields

Field	Type	Notes
prompt	list of {role, content}	Chat-format user turn
label	string	Reference answer for mathematics; empty for code
domain	string	math or code
source	string	e.g. dapo-17k, taco:codeforces
metadata	dict	rm_type (dapo / code_sandbox); code rows carry tests
difficulty	string	Code rows: EASY / MEDIUM / HARD
idx, has_starter	int, bool	Function-signature code rows

from datasets import load_dataset
ds = load_dataset("K1zE/BPM", split="train")

Redistributed for research reproduction; items remain subject to their original sources' terms.

Citation

@misc{wang2026bpm,
  title={Cross-Tokenizer On-Policy Distillation via Byte-Prefix Marginalization},
  author={Hao Wang and Kun Yuan and Wenlin Zhong and Minglei Zhang and Han Xiao and Ming Sun and Honggang Qi},
  year={2026},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  eprint={2607.22334},
}