daVinci-LLM:Towards the Science of Pretraining

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Qin, Yiwei, Liu, Yixiu, Mi, Tiantian, Xie, Muhang, Huang, Zhen, Si, Weiye, Lu, Pengrui, Feng, Siyuan, Wu, Xia, Liu, Liming, Luo, Ye, Hou, Jinlong, Guo, Qipeng, Qiao, Yu, Liu, Pengfei
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911550081597440
author Qin, Yiwei
Liu, Yixiu
Mi, Tiantian
Xie, Muhang
Huang, Zhen
Si, Weiye
Lu, Pengrui
Feng, Siyuan
Wu, Xia
Liu, Liming
Luo, Ye
Hou, Jinlong
Guo, Qipeng
Qiao, Yu
Liu, Pengfei
author_facet Qin, Yiwei
Liu, Yixiu
Mi, Tiantian
Xie, Muhang
Huang, Zhen
Si, Weiye
Lu, Pengrui
Feng, Siyuan
Wu, Xia
Liu, Liming
Luo, Ye
Hou, Jinlong
Guo, Qipeng
Qiao, Yu
Liu, Pengfei
contents The foundational pretraining phase determines a model's capability ceiling, as post-training struggles to overcome capability foundations established during pretraining, yet it remains critically under-explored. This stems from a structural paradox: organizations with computational resources operate under commercial pressures that inhibit transparent disclosure, while academic institutions possess research freedom but lack pretraining-scale computational resources. daVinci-LLM occupies this unexplored intersection, combining industrial-scale resources with full research freedom to advance the science of pretraining. We adopt a fully-open paradigm that treats openness as scientific methodology, releasing complete data processing pipelines, full training processes, and systematic exploration results. Recognizing that the field lacks systematic methodology for data processing, we employ the Data Darwinism framework, a principled L0-L9 taxonomy from filtering to synthesis. We train a 3B-parameter model from random initialization across 8T tokens using a two-stage adaptive curriculum that progressively shifts from foundational capabilities to reasoning-intensive enhancement. Through 200+ controlled ablations, we establish that: processing depth systematically enhances capabilities, establishing it as a critical dimension alongside volume scaling; different domains exhibit distinct saturation dynamics, necessitating adaptive strategies from proportion adjustments to format shifts; compositional balance enables targeted intensification while preventing performance collapse; how evaluation protocol choices shape our understanding of pretraining progress. By releasing the complete exploration process, we enable the community to build upon our findings and systematic methodologies to form accumulative scientific knowledge in pretraining.
format Preprint
id arxiv_https___arxiv_org_abs_2603_27164
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle daVinci-LLM:Towards the Science of Pretraining
Qin, Yiwei
Liu, Yixiu
Mi, Tiantian
Xie, Muhang
Huang, Zhen
Si, Weiye
Lu, Pengrui
Feng, Siyuan
Wu, Xia
Liu, Liming
Luo, Ye
Hou, Jinlong
Guo, Qipeng
Qiao, Yu
Liu, Pengfei
Artificial Intelligence
Computation and Language
The foundational pretraining phase determines a model's capability ceiling, as post-training struggles to overcome capability foundations established during pretraining, yet it remains critically under-explored. This stems from a structural paradox: organizations with computational resources operate under commercial pressures that inhibit transparent disclosure, while academic institutions possess research freedom but lack pretraining-scale computational resources. daVinci-LLM occupies this unexplored intersection, combining industrial-scale resources with full research freedom to advance the science of pretraining. We adopt a fully-open paradigm that treats openness as scientific methodology, releasing complete data processing pipelines, full training processes, and systematic exploration results. Recognizing that the field lacks systematic methodology for data processing, we employ the Data Darwinism framework, a principled L0-L9 taxonomy from filtering to synthesis. We train a 3B-parameter model from random initialization across 8T tokens using a two-stage adaptive curriculum that progressively shifts from foundational capabilities to reasoning-intensive enhancement. Through 200+ controlled ablations, we establish that: processing depth systematically enhances capabilities, establishing it as a critical dimension alongside volume scaling; different domains exhibit distinct saturation dynamics, necessitating adaptive strategies from proportion adjustments to format shifts; compositional balance enables targeted intensification while preventing performance collapse; how evaluation protocol choices shape our understanding of pretraining progress. By releasing the complete exploration process, we enable the community to build upon our findings and systematic methodologies to form accumulative scientific knowledge in pretraining.
title daVinci-LLM:Towards the Science of Pretraining
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2603.27164