Fine-tune of Qwen3-8B for the You Are a Bug improvisational dungeon master, built for the Hugging Face Build Small Hackathon "Thousand Token Wood" track.

Status: Active checkpoint. Replaced the earlier Llama-3.2-3B base after playtesting showed the 3B model wasn't holding up.

What it does

Given a structured "card" (PLAYER / THE WORLD / WHAT JUST HAPPENED / PLAYER'S TURN), outputs a single JSON object: either a say (narration) or a roll (stat check with difficulty band). Code owns all randomness, arithmetic, and sheet mutations — the model only narrates, decides if a roll is needed, and proposes the difficulty and which stat is tested.

Intended use

🪲 Drives the Gradio app here! 🪲 Very structured output tailor-made for the game engine — not for general-purpose chatting.

Training

Base: Qwen/Qwen3-8B
Method: SFT to create a LoRA, merged, quantized, and exported as GGUF (Q8_0).
Dataset: a blend of hand-curated and bulk-generated turns covering a wide range of game states, user inputs, and bug types.
Hardware: Modal.com

Limitations

JSON output is strict — the runtime has a lenient parser that degrades malformed output to say.
English only.

jnalv/you-are-a-bug-qwen3-8B

You Are a Bug — DM Model (Qwen3-8B)

What it does

Intended use

Training

Limitations

jnalv/
you-are-a-bug-qwen3-8B