SpeakSpark — SparkySpeaks-qwen3-1.7B

Kids English guessing game AI fine-tuned from Qwen3-1.7B.

Training stages

Stage	Goal
A	Kid-safe English conversation (500+ synthetic + UltraChat)
B	Grammar recast — clean corrected restatements
C	Three personalities: L1 energetic, L2 witty, L3 calm/confident
D	Magic Box JSON controller — structured {say, recast, next} output

Files

• gguf/ — q4_k_m model (~1.1 GB) for Android / llama.cpp
• adapter_stage_d/ — final LoRA adapter

Prompt format

ChatML. Use controller system prompt + difficulty level (L1/L2/L3).