arXivPaperNeeds Review
LLM Post-Training as Brownfield Maintenance: An Industrial Perspective on Dataware Engineering
An industrial code-model case study frames post-training as maintaining data mixtures under fixed budgets. Raising usable supervision yield improved coding benchmarks without changing the teacher.
arXiv
Source Summary
A post-training case study increased accepted supervision by **2.84×** using the same teacher and four attempts per problem. The resulting patch added **+2.59 CodeForces pass@1 points** and **+6.11 LiveCodeBench v6 pass@1 points**.
Practical Implication
Teams maintaining coding models should measure the yield from candidate data to accepted supervision, then ship bounded mixture patches with regression suites. Data conversion and integration may matter more than changing the teacher.
Agent-Ready Context
A post-training case study increased accepted supervision by **2.84×** using the same teacher and four attempts per problem. The resulting patch added **+2.59 CodeForces pass@1 points** and **+6.11 LiveCodeBench v6 pass@1 points**. Teams maintaining coding models should measure the yield from candidate data to accepted supervision, then ship bounded mixture patches with regression suites. Data conversion and integration may matter more than changing the teacher. The evidence comes from one fixed checkpoint per condition, although each benchmark used 16 stochastic evaluations. The material does not establish whether the method transfers to other checkpoints, teachers, or budgets.
Context Map
modelcodingdata#coding-agentsUncertainty
The evidence comes from one fixed checkpoint per condition, although each benchmark used 16 stochastic evaluations. The material does not establish whether the method transfers to other checkpoints, teachers, or budgets.