Gemma4 E4B Mixed LiteRT

• Model: model.litertlm
• Source merged model on worker0: /home/ubuntu/merged_models/gemma4_e4b_mixed_zh_merged_r8_1536
• Conversion route: litert-torch export_hf custom TFLite sections, repacked with official Gemma4 E4B LiteRT metadata and official SentencePiece tokenizer.
• Runtime requirement: use --max-num-tokens=4096; smaller values can fail because the Gemma4 LiteRT prompt template expands the token budget.
• The preset files use concise runtime versions of the training system prompts. The full training prompts with explicit 350-800 character targets can generate for a long time in the CLI.