fix: rename mcp/ to mcp_servers/ to resolve namespace collision (#3)

Add MiniAgent-494M (Qwen2.5 multilingual base), update README with all 3 models

Add continuous learning, RAG, blog training data, Qwen2.5-0.5B support

Update README with real trained models, training results, and honest quality assessment

Fix combined mode OOM: reduce batch_size to 8 and max_length to 320 for 104M model on 20GB

Add combined training mode: minimind base + advertising pretrain + SFT

Add practitioner expertise from skills, hub, and agent specs to training data

Add 60+ platform training data: Search, Social, DSP, Retail, CTV, Mobile, Native, Analytics

Add GGUF converter, update README with HuggingFace model link

Fix SFT OOM: reduce batch_size to 8 and max_length to 512 for 20GB GPUs

Move import math to top of pretrain.py (was only inside main())

Rewrite GQAttention.forward with consistent dimension ordering

Fix RoPE apply_rope: index by seq_len (dim 2) not n_heads (dim 1)

Fix device detection: send GPU info to stderr, only return device string

Fix training pipeline: real tokenizer, expanded data, one-command GPU training