SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.10460 · ACL · 2025

Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond

Xin He, Yongchao Deng, Liang Wen, Haosheng Zou, Xiangzheng Zhang, Qi An, Yunke Cai, Fenrui Xiao, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, and 3 more

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 13 functions out of this paper's own repositories and ran 9 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
Qihoo360/Light-R1 canonical 9 of 13
FunctionStatusWhere it lives
build_ngram_lookup Ran Qihoo360/Light-R1/decontaminate/n_gram_check.py
code served (permissive licence) · get_code("57f0335f22430956")
clean_text Ran Qihoo360/Light-R1/decontaminate/character_matching.py
code served (permissive licence) · get_code("61619ba901c7b574")
count_unknown_letters_in_expr Ran Qihoo360/Light-R1/deepscaler-release/deepscaler/rewards/math_utils/utils.py
code served (permissive licence) · get_code("368d3db6347269c3")
create_huggingface_actor Ran Qihoo360/Light-R1/deepscaler-release/verl/verl/utils/model.py
code served (permissive licence) · get_code("360c671d6041ba3b")
normalize_string Ran Qihoo360/Light-R1/decontaminate/n_gram_check.py
code served (permissive licence) · get_code("927fae1304011f13")
should_allow_eval Ran Qihoo360/Light-R1/deepscaler-release/deepscaler/rewards/math_utils/utils.py
code served (permissive licence) · get_code("95f02454eddf082a")
union_tensor_dict Ran Qihoo360/Light-R1/deepscaler-release/verl/verl/protocol.py
code served (permissive licence) · get_code("21331a58f93375e4")
unpad_dataproto Ran Qihoo360/Light-R1/deepscaler-release/verl/verl/protocol.py
code served (permissive licence) · get_code("25f0ea3f460f6ce1")
word_ngrams Ran Qihoo360/Light-R1/decontaminate/n_gram_check.py
code served (permissive licence) · get_code("b740be8e99a575cb")
get_huggingface_actor_config Not yet run Qihoo360/Light-R1/deepscaler-release/verl/verl/utils/model.py
code served (permissive licence) · get_code("19472d4343d0a082")
get_weight_loader Not yet run Qihoo360/Light-R1/deepscaler-release/verl/verl/models/weight_loader_registry.py
code served (permissive licence) · get_code("919cf310ad4c189a")
mathd_normalize_answer Not yet run Qihoo360/Light-R1/deepscaler-release/deepscaler/rewards/math_utils/utils.py
code served (permissive licence) · get_code("ec578739f88adc22")
squeeze Not yet run Qihoo360/Light-R1/deepscaler-release/verl/verl/utils/model.py
code served (permissive licence) · get_code("3b15e2ac7497c441")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

This paper introduces Light-R1, an opensource suite for training long reasoning models using reproducible and cost-effective methodology. Given the proprietary nature of data used in the DeepSeek-R1 series, we develop an alternative approach leveraging exclusively public data and models. Our curriculum training progressively increases data difficulty, combined with multi-staged post-training. Our Light-R1-32B model, trained from Qwen2.5-32B-Instruct, outperforms DeepSeek-R1-Distill-Qwen-32B in math reasoning. Experimental results show that this curriculum approach becomes more effective when distinct, diverse datasets are available for different training stages: fine-tuning DeepSeek-R1-Distilled models (pre-tuned by DeepSeek team on proprietary data) with 3,000 challenging examples from our curriculum dataset yielded state-ofthe-art 7B and 14B models, while the 32B model, Light-R1-32B-DS performed comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our work by applying GRPO on long reasoning models. Our final Light-R1-14B-DS achieves SOTA performance among 14B models in math, with AIME24 & 25 scores of 74.0 and 60.2 respectively, surpassing many 32B models and DeepSeek-R1-Distill-Llama-70B. Despite math-focused training, Light-R1-14B-DS demonstrates strong cross-domain generalization. Light-R1 represents a significant advancement in making sophisticated reasoning models more accessible and implementable in real-world applications. Our models, training data and code have been made available at https://github.com/Qihoo360/Light-R1.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2503.10460")
get_code_for_paper("2503.10460")
have("2503.10460")

Connect an agent — have() is free.