本项目从零开始构建并优化了一个千万参数级别的大规模预训练语言模型，涵盖预训练、有监督微调（SFT）和R1推理蒸馏三个阶段。项目采用自定义Transformer架构（包括RMSNorm、分组注意力、多Query机制、SwiGLU激活和RoPE位置编码），实现高效的长文本处理和自回归生成。同时，开发了基于BBPE编码的分词器，针对中文进行了优化，构建了轻量级词表。预训练阶段通过混合精度训练、梯度累积和Cosine Annealing学习率调度，实现全流程训练；微调阶段则利用开源SFT数据集，通过特殊损失掩码优化指令遵循；R1推理蒸馏阶段采用Deepseek-R1数据及损失加权策略，赋予模型逐步推理（慢思考）的能力，显著提升复杂任务的回答质量。
☆23Mar 10, 2025Updated last year
Alternatives and similar repositories for WeakWater-30M
Users that are interested in WeakWater-30M are comparing it to the libraries listed below. We may earn a commission when you buy through links labeled 'Ad' on this page.
Sorting:
- ☆13Aug 19, 2025Updated 11 months ago
- 使用Decoder-only的Transformer进行时序预测，包含SwiGLU和RoPE(Rotary Positional Embedding)，Time series prediction using Decoder-only Transformer, Includ…☆16Jan 25, 2024Updated 2 years ago
- [KDD 2026] Official implementation of "FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via…☆17Jun 1, 2026Updated 2 months ago
- FastTrack4LLM 是一个为大模型学习者准备的大模型学习与实践框架，帮助他们轻松掌握大模型的核心原理与训练流程，让每个人都能真正理解大模型的内部机制。本项目不仅完整复现了 LLaMA、Qwen、DeepSeek 等主流开源大模型架构，还覆盖了大模型的全生命周期：To…☆31Nov 6, 2025Updated 9 months ago
- hrtem lattice fringe image analysis algorithms - developed for carbon nanostructure analysis☆10Jan 9, 2026Updated 7 months ago
- Virtual machines for every use case on DigitalOcean • AdGet dependable uptime with 99.99% SLA, simple security tools, and predictable monthly pricing with DigitalOcean's virtual machines, called Droplets.
- An Efficent BPE Algorithm Faster then Hugging Face Tokenizer's Implementation☆13Sep 9, 2024Updated last year
- 北京邮电大学网络工程嵌入式系统实验报告☆12Jan 7, 2021Updated 5 years ago
- 最基本最小白的自然语言处理入门读物，基于deepseek-r1，涵盖了传统NLP和现代大模型☆30Jan 16, 2026Updated 6 months ago
- 一个基于transformers的自定义命名实体识别模型示例☆17Jul 4, 2021Updated 5 years ago
- ☆18Feb 6, 2025Updated last year
- 校园音乐征集投票系统 A system for electing annual school music☆10Jul 31, 2026Updated 2 weeks ago
- A comprehensive evaluation framework for Backboard's memory system using the LoCoMo (Long Conversation Memory) benchmark.☆17Nov 3, 2025Updated 9 months ago
- 一个基于 模型上下文协议/MCP 构建的智能医学文献分析工具。它旨在帮助科研人员、医学从业者和学生快速检索 PubMed 数据库，并利用大型语言模型 (LLM) 的能力对文献摘要进行智能分析和总结☆11Jun 17, 2026Updated last month
- Official implementation for "Enhancing Semantics in Multimodal Chain of Thought via Soft Negative Sampling"☆10May 21, 2024Updated 2 years ago
- Open source password manager - Proton Pass • AdSecurely store, share, and autofill your credentials with Proton Pass, the end-to-end encrypted password manager trusted by millions.
- This repository includes code and materials for the paper "Efficient PRM Training Data Synthesis via Formal Verification" (ACL 2026 Findi…☆19Apr 7, 2026Updated 4 months ago
- 使用分布式技术结合前端Vue框架，后端mycat+mybatis+springboot框架实现金融分布式交易系统，包括：开户，登录，信息查询，明细查询，一对一转账，计息，对账等金融交易的核心功能。☆12Sep 3, 2020Updated 5 years ago
- 反洗钱使用黑名单数据爬取☆14Jul 7, 2015Updated 11 years ago
- 手撕transformer并完成一个简单的机器翻译。☆22Feb 19, 2025Updated last year
- ☆15Aug 31, 2025Updated 11 months ago
- ETL调度管理平台☆12Nov 21, 2022Updated 3 years ago
- A Real-Time Network Intrusion Detection System based on Transformer-LSTM model (Major Project)☆16May 27, 2024Updated 2 years ago
- 使用Sentencepiece对中文语料进行分词☆13Nov 30, 2023Updated 2 years ago
- EasyTTS是一个便捷的工具，旨在方便地使用第三方API服务来调用OpenAI的文本转语音（TTS）功能。 EasyTTS允许用户输入文本，并选择不同的模型、音色、格式来生成音频文件。☆10Nov 26, 2023Updated 2 years ago
- Deploy on Railway without the complexity - Free Credits Offer • AdConnect your repo and Railway handles the rest with instant previews. Quickly provision container image services, databases, and storage volumes.
- A python script that you can execute form terminal to directly download a paper that you want form arXiv☆17Jul 3, 2017Updated 9 years ago
- 「城语」APP基于A级景区、历史古迹、文物保护单位等基础数据，利用先进的大模型能力实现智能化的Citywalk 路线规划，包括设计一条路线、生成路线攻略、生成景点的推荐理由等三大核心功能；利用大模型减少了人工编辑和推荐的工作量，并可以根据游客的需求进行个性化定制，提升了游客…☆19Feb 20, 2024Updated 2 years ago
- Chinese Word Segmentation task based on BERT and implemented in Pytorch☆14Aug 14, 2020Updated 6 years ago
- 前后端分离的答题系统，前端对应pc的网页，和对应微信的小程序端，前期主要对微信小程序进行完善。前端网页设计使用vue技术，后端使用SpringBoot等相关技术☆15Aug 18, 2020Updated 5 years ago
- How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients☆20Jun 17, 2025Updated last year
- Classification and aggregation of russian news articles. University coursework.☆18Jan 21, 2019Updated 7 years ago
- Identify the type of news based on headlines and short descriptions☆16Mar 25, 2019Updated 7 years ago
- 本项目利用医学领域的 CoT 数据对 Deepseek-R1-Distill-Qwen-7B 进行微调，通过 QLoRA 量化和 Unsloth 加速训练，显著提升模型在复杂医学推理任务中的慢思考能力。知识蒸馏技术使轻量级模型获得大模型的推理优势，实现高效、准确且具有解释性…☆48Mar 10, 2025Updated last year
- Code for Predictive Network Representation Learning (SIGIR 2017)☆11Apr 24, 2020Updated 6 years ago
- End-to-end encrypted cloud storage - Proton Drive • AdSpecial offer: 40% Off Yearly / 80% Off First Month. Protect your most important files, photos, and documents from prying eyes.
- ☆17Dec 16, 2015Updated 10 years ago
- ☆15Oct 19, 2025Updated 9 months ago
- 基于CNN的新浪新闻文本分类☆11Jul 22, 2019Updated 7 years ago
- A Tight-fisted Optimizer (Tiger), implemented in PyTorch.☆12Jun 26, 2024Updated 2 years ago
- Testing the performance of CNN and BERT embeddings on GLUE tasks☆15Mar 24, 2023Updated 3 years ago
- ☆16Jul 17, 2025Updated last year
- Partitioning a large Embedding LUT into multiple devices. (for Pytorch)☆16Apr 30, 2017Updated 9 years ago