LumiVore/lumivore-stage2-training-data
收藏Hugging Face2026-03-22 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/LumiVore/lumivore-stage2-training-data
下载链接
链接失效反馈官方服务：
资源简介：
# Lumivore Stage 2 Training Dataset
**Version:** 4.5-B
**Created:** March 2026
**Purpose:** LoRA fine-tuning on OpenClaw-specific agent behavior
**Total Examples:** ~11,880
**Format:** Alpaca (instruction, input, output)
---
## Overview
This dataset was used for Stage 2 of the Lumivore-1.2B training pipeline. It focuses specifically on OpenClaw agent behavior — tool use, structured reasoning, and agentic task execution patterns.
Stage 2 was a LoRA fine-tuning (rank=64) on top of the Stage 1 full fine-tuning, allowing the model to specialize while preserving general capabilities.
---
## Data Sources
| Source | Description | Proportion |
|--------|-------------|------------|
| **OpenClaw v3.0** | Session logs with tool reasoning chains | ~50% |
| **Agent Tool Use** | OpenClaw docs-derived examples | ~25% |
| **ClawHub Skills** | Skill documentation and usage patterns | ~25% |
---
## Dataset Characteristics
- **Task types:** Tool calls, agent loops, reasoning chains, structured JSON outputs
- **Style:** Agentic, action-oriented, OpenClaw-specific
- **Quality:** Filtered for correctness, PII sanitized
- **Augmentation:** 5x linguistic variations
---
## Training Configuration
Used with the following hyperparameters:
```python
# Stage 2 Training (LoRA)
- Base model: Stage 1 output (/workspace/checkpoints/stage1/final)
- LoRA rank: 64, alpha: 128
- LoRA targets: q_proj, v_proj, gate
- Batch size: 1 (micro), gradient_accumulation: 16
- Effective batch: 16
- Max sequence length: 1024
- Learning rate: 5e-6 (10x lower than Stage 1)
- Optimizer: 8-bit AdamW
- Epochs: 3
- Steps: 2,115
- Duration: ~5 hours on AMD RX 7600 XT
- Peak VRAM: ~3.3/16.0 GB
```
---
## Files
- `train.jsonl` — Training examples (~11,288 after split)
- `validation.jsonl` — Validation examples (~5% split)
- `metadata.json` — Dataset metadata and statistics
- `README.md` — This documentation
---
## Usage
```python
from datasets import load_dataset
dataset = load_dataset("LumiVore/lumivore-stage2-training-data")
train_data = dataset["train"]
val_data = dataset["validation"]
```
---
## Related
- **Stage 1 Dataset:** `LumiVore/lumivore-stage1-training-data` — General agentic fine-tuning
- **Stage 3 Dataset:** `LumiVore/lumivore-stage3-identity-dataset` — Identity and conversational training
- **Model:** `LumiVore/lumivore-1.2b` (when published)
---
## Citation
If you use this dataset, please cite:
```bibtex
@dataset{lumivore2026stage2,
title={Lumivore Stage 2 Training Dataset},
author={LumiVore AI},
year={2026},
url={https://huggingface.co/datasets/LumiVore/lumivore-stage2-training-data}
}
```
---
*Created for the Lumivore-1.2B training pipeline*
# Lumivore 第二阶段训练数据集
**版本：** 4.5-B
**创建时间：** 2026年3月
**用途：** 针对OpenClaw特定智能体行为的低秩适配（Low-Rank Adaptation, LoRA）微调
**总样本数：** 约11880条
**格式：** Alpaca格式，字段包含instruction（指令）、input（输入）、output（输出）
---
## 概述
本数据集用于Lumivore-1.2B训练流水线的第二阶段，专门聚焦OpenClaw智能体行为——涵盖工具调用、结构化推理以及智能体任务执行模式。
第二阶段是在第一阶段全量微调的基础上开展低秩适配（LoRA）微调（秩=64），可在保留模型通用能力的同时实现专业化适配。
---
## 数据源
| 数据源 | 描述 | 占比 |
|--------|-------------|------------|
| **OpenClaw v3.0** | 包含工具推理链的会话日志 | 约50% |
| **Agent Tool Use** | 源自OpenClaw官方文档的示例数据 | 约25% |
| **ClawHub Skills** | 技能文档与实际使用模式 | 约25% |
---
## 数据集特征
- **任务类型：** 工具调用、智能体循环、推理链、结构化JSON输出
- **风格：** 智能体导向、行动取向、专为OpenClaw定制
- **质量：** 经过正确性过滤，已完成个人可识别信息（Personally Identifiable Information, PII）脱敏处理
- **数据增强：** 进行了5倍的语言变体扩充
---
## 训练配置
使用以下超参数进行训练：
python
# 第二阶段LoRA训练
- 基础模型：第一阶段最终输出（/workspace/checkpoints/stage1/final）
- LoRA秩：64，alpha值：128
- LoRA目标模块：q_proj、v_proj、gate
- 微批次大小：1，梯度累积步数：16
- 有效批次大小：16
- 最大序列长度：1024
- 学习率：5e-6（比第一阶段低10倍）
- 优化器：8-bit AdamW
- 训练轮次：3
- 训练步数：2115
- 训练时长：在AMD RX 7600 XT显卡上约5小时
- 峰值显存占用：约3.3/16.0 GB
---
## 数据集文件
- `train.jsonl` — 训练样本集（拆分后约11288条）
- `validation.jsonl` — 验证样本集（占拆分集的约5%）
- `metadata.json` — 数据集元数据与统计信息
- `README.md` — 本文档说明
---
## 使用方法
python
from datasets import load_dataset
dataset = load_dataset("LumiVore/lumivore-stage2-training-data")
train_data = dataset["train"]
val_data = dataset["validation"]
---
## 相关资源
- **第一阶段数据集：** `LumiVore/lumivore-stage1-training-data` — 通用智能体微调数据集
- **第三阶段数据集：** `LumiVore/lumivore-stage3-identity-dataset` — 身份与对话训练数据集
- **配套模型：** `LumiVore/lumivore-1.2b`（待正式发布）
---
## 引用规范
若使用本数据集，请引用如下文献：
bibtex
@dataset{lumivore2026stage2,
title={Lumivore Stage 2 Training Dataset},
author={LumiVore AI},
year={2026},
url={https://huggingface.co/datasets/LumiVore/lumivore-stage2-training-data}
}
---
*专为Lumivore-1.2B训练流水线打造*
提供机构：
LumiVore
