Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Building high-quality Q&A data for Qwen3-32B LoRA fine-tuning

AI summary

This tutorial explains how to construct question-and-answer datasets for fine-tuning a Qwen3-32B model that previously underwent LoRA incremental pretraining. It covers manual annotation, knowledge-base-assisted generation, model distillation, structured-data conversion, real-world data回流, and quality control covering deduplication, model scoring, and contamination checks.

In-depth insight

The article provides a concrete workflow for turning source documents and operational data into traceable, high-quality SFT examples, including code for RAGFlow-backed answers, voting-based distillation, and dataset filtering.

前言

上篇文章大模型训练全流程实战指南实战篇(十七)—— LlamaFactory 预训练实战:从语料到模型中,笔者使用76457条、约926M tokens的预训练数据集,完成了Qwen3-32B的LoRA增量预训练,并选定checkpoint-6250作为下一阶段的基座模型。

经过增量预训练后的基座模型有一个典型特点:面对用户提问时,它的回答更像“顺着文档继续写”,而不是“精准回答用户的问题”。例如,当用户提出:

我有一台对外暴露的 Redis 6.2,怀疑存在未授权访问。请告诉我怎么确认、具体改哪些配置、改完怎么验证。

增量预训练后的模型可能给出如下回复:

Redis 是一个广泛使用的高性能内存数据库,常用于缓存、消息队列和会话存储。由于默认配置较为宽松,
如果 Redis 直接暴露在公网,容易发生未授权访问漏洞。攻击者可以利用未授权访问写入恶意文件、植入
挖矿程序、窃取数据,甚至进一步控制服务器。建议用户设置强密码、绑定内网地址、禁用危险命令、
配置防火墙,并定期进行安全巡检。同时应关注 Redis 官方安全公告,及时升级到最新版本。

#Qwen3-32B#LLaMA-Factory#RAGFlow#Easy Dataset#Qwen-plus

Original · 21 min readRead the original
Group chat

No comments yet. Start the conversation.