solana-clawd-instruct
收藏资源简介:
Solana Clawd Instruct 是一个精心策划的指令微调数据集,专门用于将基础语言模型微调为具有Solana原生能力的Clawd智能体。该数据集广泛覆盖Solana区块链技术(如PDA、账户、指令、租金、计算预算、Token-2022)、去中心化金融(DeFi)核心概念(如AMM、CLMM、永续合约、绑定曲线、Jupiter、Phoenix)、Memecoin风险分析(如跑路检测、持有者集中度、部署者取证)、智能体架构(如技能注册表、脑/手分离、多智能体协调)、宪法推理(Clawd宪法、安全护栏、拒绝模式)、代码生成(Anchor/Rust、TypeScript @solana/kit、Python)以及零知识证明(ZK)压缩(Light Protocol、nullifiers、Groth16)等多个关键领域。数据集采用OpenAI消息对话格式,每个样本为一个包含系统提示、用户查询和助手回复的完整对话。系统提示在所有样本中保持一致,以确保微调后的模型能锁定Clawd的对话风格和宪法安全护栏。数据来源于solana-clawd仓库的官方文档、公开的Solana/DeFi参考资料、社区最佳实践清单以及针对边缘案例合成的宪法场景。数据集适用于微调基础指令模型、进行Solana领域的持续预训练、训练工具调用能力(如Hermes-3路径的13种工具套件),以及作为评估模型在Solana知识准确性和宪法对齐方面的基准。数据集不包含任何实时交易数据、抢先交易示例、制裁规避、KYC绕过或钱包盗取等有害内容,并明确拒绝协助此类行为。数据集遵循CC-BY-4.0许可证。
Solana Clawd Instruct is a meticulously curated instruction fine-tuning dataset specifically developed to fine-tune base language models into Clawd AI Agents equipped with natively Solana-integrated capabilities. It extensively covers multiple critical domains, including Solana blockchain technologies (e.g., PDA, accounts, instructions, rent, compute budget, Token-2022), core decentralized finance (DeFi) concepts (e.g., AMM, CLMM, perpetual contracts, bonding curves, Jupiter, Phoenix), memecoin risk analysis (e.g., rug pull detection, holder concentration, deployer forensics), agent architectures (e.g., skill registry, brain-hand separation, multi-agent coordination), constitutional reasoning (Clawd Constitution, safety guardrails, refusal modes), code generation (Anchor/Rust, TypeScript @solana/kit, Python), and zero-knowledge proof (ZK) compression (Light Protocol, nullifiers, Groth16). The dataset follows the OpenAI message conversation format, where each sample constitutes a complete dialogue containing a system prompt, user query, and assistant reply. The system prompt is consistent across all samples, ensuring that the fine-tuned model adheres to Clawd's dialogue style and constitutional safety guardrails. The dataset is derived from official documentation in the solana-clawd repository, publicly available Solana/DeFi reference materials, community-curated best practice lists, and constitutional scenarios synthesized for edge cases. This dataset is suitable for fine-tuning base instruction models, conducting continued pre-training in the Solana domain, training tool-calling abilities (such as the 13-tool suite of the Hermes-3 pathway), and acting as a benchmark for evaluating model accuracy in Solana knowledge and constitutional alignment. The dataset contains no harmful content such as real-time trading data, front-running examples, sanction evasion, KYC bypass, or wallet theft, and explicitly prohibits assistance with such activities. The dataset is released under the CC-BY-4.0 license.
Solana Clawd Instruct 数据集概述
数据集名称:Solana Clawd Instruct
许可协议:CC-BY-4.0
语言:英语
任务类别:文本生成、问答
标签:Solana、DeFi、加密货币、Agent、Clawd、Constitutional AI
数据规模:少于1000条样本
数据集大小:约44.59 MB(下载大小44.56 MB)
数据格式
每条样本为OpenAI messages 格式的单一对话记录,包含以下字段:
role:字符串,角色(system、user、assistant)content:字符串,对话内容
示例结构: json { "messages": [ {"role": "system", "content": "You are Clawd, a sovereign Solana-native AI agent..."}, {"role": "user", "content": "What is a PDA?"}, {"role": "assistant", "content": "A PDA is a Program Derived Address..."} ] }
系统提示在所有样本中保持一致,以固定模型的Clawd语音和Constitutional防护。
数据划分
数据集按90/5/5比例划分(种子42),具体如下:
| 划分 | 样本数 | 用途 |
|---|---|---|
train |
42 | SFT训练 |
eval |
2 | 训练时验证损失 |
test |
3 | 独立保留评估 |
划分是确定性的(种子42),另有独立保留文件 data/solana_clawd_eval.jsonl 从未用于训练。
教学内容
数据集覆盖以下领域,用于将模型微调为Solana原生Clawd Agent:
- [x] Solana机制(PDA、账户、指令、租金、计算预算、Token-2022)
- [x] DeFi原语(AMM、CLMM、永续合约、债券曲线、Jupiter、Phoenix)
- [x] Memecoin风险分析(Rug检测、持仓集中度、部署者取证)
- [x] Agent架构(技能注册、大脑/手分工、多Agent协调)
- [x] Constitutional推理(Clawd宪法、防护栏、拒绝模式)
- [x] 代码生成(Anchor/Rust、TypeScript @solana/kit、Python)
- [x] ZK压缩(Light Protocol、nullifier、Groth16)
- [ ] Perps函数调用(13个工具——Phoenix、Jupiter、风险评估)
- [ ] Runtime v2(xAI语音Agent、MCP技能目录、ClawdRouter、x402)
数据来源
solana-clawd仓库文档(AGENTS.md、CONSTITUTION.md、skills/)- 公开Solana/DeFi参考资料(Anchor文档、Helius SDK、Jupiter、Phoenix)
- 最佳实践的Memecoin风险检查清单
- Constitutional边缘场景的合成示例
- 本地ZK参考和示例(来源于
zk-primitives/)
所有数据均为原创、源自公开文档或常见模式的清晰重述,不包含专有策略代码。
预期用途
- 将基础指令模型(Qwen2.5-1.5B、Hermes-3-8B、Llama-3.2-1B)微调为Clawd语音
- 继续预训练特定领域模型
- Hermes-3路径的工具使用训练(13个工具的Perps函数调用套件)
- Solana准确性和Constitutional对齐的评估基准
不适用范围
- 实时交易数据:不包含真实钱包交易、盈亏数据或账户余额
- 抢跑示例:数据集有意规避进攻性MEV
- 制裁规避、KYC绕过、钱包盗取:系统提示中已拒绝,且所有助手回复中不存在
- 保证财务结果:本数据集为教育目的,非交易信号
许可与引用
- 许可:CC-BY-4.0,允许使用、修改和再分发(需注明出处)
- 引用格式: bibtex @misc{solana-clawd-instruct-2026, title = {Solana Clawd Instruct}, author = {solanaclawd}, year = {2026}, url = {https://huggingface.co/datasets/solanaclawd/solana-clawd-instruct} }





