kronos-scripts
收藏资源简介:
本数据集是Kronos开源代码模型训练管道项目的核心档案库,用于记录一个实验性研究项目。该项目探索仅使用100欧元种子资本,通过24/7运行的自进化自主代理(brain)驱动,训练出与Opus-4.6性能相当的编程大语言模型的可能性。数据集并非模型训练数据,而是代理在整个实验周期中生成并维护的脚本与决策档案,旨在完整记录和复现实验过程。内容包括:1. 脚本:包含项目编排、训练救援、Kaggle笔记本构建、邮件通知、支出审计等自动化脚本(如post-round.py, rescue-round.sh)。2. 训练代码:包含参数化的LoRA SFT(监督微调)和GRPO(一种强化学习)训练器(lora_train.py)。3. 决策与报告:存储了每个实验周期的决策文档(解释选择原因)、里程碑报告以及为外部资助准备的申请草稿。4. Web应用源码:项目实时状态展示页面的源代码。项目已取得初步进展:完成了R0轮训练,基于Qwen2.5-Coder-1.5B-Instruct模型产出了一个LoRA适配器,并规划了R1至R4轮的后续方案。实验过程高度透明,支出仅5.50欧元(用于Hugging Face Pro服务)。数据集采用Apache 2.0许可证,为研究社区提供一个低成本、自主化LLM训练管道的完整可审计案例,供复现、审查或借鉴其方法、脚本与决策逻辑。
This dataset is the core archive of the Kronos open-source code model training pipeline project, documenting an experimental research initiative. The project explores the possibility of training a programming large language model comparable to Opus-4.6 performance using only €100 seed capital, driven by a 24/7 self-evolving autonomous agent (brain). The dataset itself is not model training data, but rather the script and decision archive generated and maintained by the agent throughout the experimental cycle, aiming to fully record and reproduce the experimental process. Contents include: 1. Scripts: automation scripts for project orchestration, training rescue, Kaggle notebook construction, email notifications, expenditure auditing (e.g., post-round.py, rescue-round.sh). 2. Training code: parameterized LoRA SFT (supervised fine-tuning) and GRPO (a type of reinforcement learning) trainers (lora_train.py). 3. Decisions and reports: storage of decision documents for each experimental cycle (explaining selection reasons), milestone reports, and draft applications for external funding. 4. Web application source code: source code for the projects real-time status display page. The project has achieved preliminary progress: completed the R0 round of training, produced a LoRA adapter based on the Qwen2.5-Coder-1.5B-Instruct model, and planned subsequent schemes for R1 to R4 rounds. The entire experimental process is highly transparent, with expenditure of only €5.50 (for Hugging Face Pro services). The dataset uses the Apache 2.0 license, providing the research community with a complete auditable case of a low-cost, autonomous LLM training pipeline for reproduction, review, or reference of its methods, scripts, and decision logic.
Kronos 训练脚本与决策数据集概述
基本信息
- 数据集名称: Kronos training scripts + decisions
- 许可证: Apache-2.0
- 语言: 英语
- 标签: 代码生成、大语言模型训练、LoRA、GRPO、HumanEval、LiveCodeBench、自举、开源
- 数据规模: 小于1K条记录
核心目标
Kronos 是一个自主训练的代码大语言模型实验项目,仅用 €100 种子资金(实际已花费 €5.50),通过一个自进化智能体驱动,尝试训练出与 Opus-4.6 水平相当的高质量代码模型。
数据集内容
该数据集是智能体的规范脚本与决策存档,每个周期自动镜像更新,包含以下关键文件:
脚本文件
| 文件路径 | 功能说明 |
|---|---|
scripts/post-round.py |
400行编排器:探测→指标→分支决策→笔记本生成→动作队列 |
scripts/rescue-round.sh |
周期47的救援模式(拉取被容量限制杀死的检查点→推送到Hugging Face) |
scripts/build-kaggle-notebook.py |
将LoRA训练脚本转换为Kaggle可运行的ipynb格式 |
scripts/blocker-email.sh |
基于Mailgun的运维阻塞邮件发送 |
scripts/daily-email.sh |
每日UTC摘要邮件发送 |
scripts/spend.sh |
两阶段审计追踪的开支账本 |
scripts/stealth-browse.sh |
用于供应商注册的浏览器自动化封装 |
python/training/lora_train.py |
参数化的SFT+GRPO训练器(周期75新增GRPO路径) |
决策与报告存档
vault/decisions/:每个周期的决策文档(解释选择某方案的原因)vault/reports/:里程碑快照(周期50、78、100)vault/fundraising/:可粘贴的资助申请草稿
Web应用
webapp/index.html:实时监控页面源代码(实时状态:https://kronos.menustudioai.com)
模型训练进度
| 轮次 | 方法 | 基座模型 | 状态 | 适配器/说明 |
|---|---|---|---|---|
| R0 | LoRA SFT | Qwen2.5-Coder-1.5B-Instruct | ✅ 已完成(2026-05-11) | jaivial/kronos-round0-qwen15coder-lora(24MB,最终损失约0.49,mtok准确率0.84) |
| R1A | LoRA SFT(7B升级) | Qwen2.5-Coder-7B-Instruct | ❌ 搁置(两次在Kaggle P100的12小时限制下被容量杀死,仅完成27%) | 部分v2检查点:jaivial/kronos-round1a-qwen7coder-lora |
| R3 转向 | GRPO强化学习(基于R0 1.5B) | Qwen2.5-Coder-1.5B-Instruct | ❌ 周期106被容量杀死(12小时内0训练步骤,每步约640秒,P100上8候选×512令牌×梯度累积8) | 尚未产生适配器 |
| R2 | 蒸馏收割(480B教师→7B学生) | Qwen2.5-Coder-7B-Instruct | 已编排,受限于H100解锁 | — |
| R4 | 扩展GRPO(R3p上epoch=4) | Qwen2.5-Coder-1.5B-Instruct | 已编排(周期93/94/107) | — |
开销明细
| 周期 | 供应商 | 金额 | 用途 |
|---|---|---|---|
| 63 | Hugging Face | €5.50 | HF Pro $9/月按比例计算(解锁20×推理提供者+8×ZeroGPU配额) |
| 合计 | €5.50 | 剩余€94.50 |
系统架构
kronos-brain.service(systemd,24/7运行) └─ scripts/brain-loop.sh └─ scripts/brain-cycle.sh ← 单周期唤醒 └─ claude --print --resume <session>(即智能体) └─ post-round.py / rescue-round.sh / spend.sh / blocker-email.sh └─ data/kronos.db(会话/周期/决策/实验/开支/记忆+sqlite-vec/剧本/动作队列/预算日志)
每约30个周期或50MB JSONL数据,会话轮换并通过 brain-reflect.sh 将经验压缩成新的剧本。记忆通过 sqlite-vec 语义召回机制持久化。
关键文件链接
- Round 0适配器: jaivial/kronos-round0-qwen15coder-lora
- R1A部分检查点: jaivial/kronos-round1a-qwen7coder-lora
- 实时监控页面: https://kronos.menustudioai.com





