【Hackathon 10th Spring No.11】TrinityLLM模型复现

[r-cloudforge]

通过迁移 https://github.com/ningliu-iga/TrinityLLM 项目实现 TrinityLLM 的PaddlePaddle版本复现。

主要改动

1. 新增 ppmat/models/trinityllm/ 模型目录 — 完整的TrinityLLM实现，包括SMILES tokenizer和材料性质预测LLM
2. 新增 property_prediction/configs/trinityllm/ 训练配置
3. 新增 test/trinityllm/ 测试代码 — 13个单元测试

模型特性

• 基于SMILES的材料大语言模型
• 自定义SMILES Tokenizer，支持分子表达式解析
• RoPE (旋转位置编码) 注意力机制
• SwiGLU 前馈网络
• RMSNorm 归一化
• 基于PaddleMaterials标准API
• 已注册到 ppmat/models/__init__.py

测试

cd test && python -m pytest trinityllm/ -v
# 13 passed

参考链接

• 原始论文: https://arxiv.org/abs/2412.11825
• 原始仓库: https://github.com/ningliu-iga/TrinityLLM (MIT License)
• Hackathon Issue: 【Hackathon 10th】开源贡献个人挑战赛 · 春节特别季 Paddle#77429

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

cloudforge1 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[paddle-bot]

Thanks for your contribution!