Add 3 elite-tier training frameworks (Run #128)

[alvinreal]

Category 7: Training & Fine-tuning Ecosystem

This PR adds 3 elite-tier training frameworks that meet all criteria (1000+ stars, active development, OSI-approved license, production-ready):

Added Projects

1. MiniMind (~12.5k stars, Apache 2.0)

   • Train a 64M-parameter LLM from scratch in 2 hours for $3
   • Complete implementation: MoE, pretraining, SFT, LoRA, RLHF (DPO/PPO/GRPO)
   • Pure PyTorch without high-level abstractions - educational focus
   • Active development with recent commits (2025)

2. FastChat (~40k stars, Apache 2.0)

   • Open platform for training, serving, and evaluating LLM chatbots
   • Powers Chatbot Arena (lmarena.ai) serving 10M+ requests
   • Training code for Vicuna, MT-Bench evaluation
   • Distributed multi-model serving with OpenAI-compatible APIs
   • Active development with continuous updates

3. PaddleNLP (~12.8k stars, Apache 2.0)

   • Baidu's powerful LLM library built on PaddlePaddle
   • Supports 100+ models with efficient training and compression
   • Features RsLoRA+ algorithm, DeepSeek V3/R1 support
   • FP8/INT8 quantization and unified checkpointing
   • Active development with frequent releases (2025)

Verification

• All projects: 1000+ GitHub stars
• All projects: Active within last 3 months
• All projects: OSI-approved licenses (Apache 2.0)
• All projects: Production-ready quality

Research Notes

• Category index: 6 (Training & Fine-tuning Ecosystem)
• Previous category: 5 (Prompt Engineering & Management)
• Total runs: 128
• Projects found: 3
• Projects added: 3

---

Automated research run #128 for awesome-opensource-ai