build-small-hackathon/tiny-quest-radio-aya-lora

Tiny Quest Radio GGUF

Tiny Quest Radio GGUF is the planned local runtime model for the Tiny Quest Radio Space. It starts from CohereLabs/tiny-aya-global, receives a small app-format SFT adapter for radio adventure sections, then is merged and quantized to GGUF for llama.cpp.

Intended Use

• Audio-first branching adventure narration
• Compact all-ages interactive fiction
• Local/offline-friendly llama.cpp runtime

Output Format

=== BROADCAST ===
=== CHOICES ===
=== INVENTORY ===
=== DANGER ===
=== RADIO CLUE ===

Training Recipe

• Base model: CohereLabs/tiny-aya-global
• Method: QLoRA SFT followed by merge and GGUF quantization
• Hardware: Modal NVIDIA A10G for adapter training
• Runtime: llama-cpp-python CPU inference from a Q4_K_M GGUF file
• Dataset: synthetic radio adventure scenes in the current production format

Limitations

The model is for playful fiction. It can lose continuity in long sessions and should not be treated as a factual assistant.