CoolFace
Apppublic

Zhenzewu/tokenwood

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes
App README

🍄 Tokenwood — a very small wood that thinks

Five tiny creatures. One tiny model. Drop a button, start a saga.

Tokenwood is a pocket terrarium-world: an anxious hedgehog economist, a haiku-only snail poet, a sensationalist owl journalist, a washed-up opera frog and a magpie entrepreneur live little autonomous lives. Each one is an agent running on a single fine-tuned Qwen3-1.7B through llama.cpp — on this Space's own CPU. No cloud APIs. No GPUs. No adult supervision.

You are the Wanderer. Leave a shiny button at the Old Oak. Tell the magpie a secret about the moon. Make it snow. Then let time pass — and read what the wood believes happened in tonight's edition of The Thousand Token Tribune, Tokenwood's only (and therefore finest) newspaper.

🎬 Demo video: link added at submission 📣 Social post: link added at submission 📓 Field notes (blog): link added at submission 🧠 Fine-tuned model: link added at publish 📡 Agent traces dataset: link added at publish

Why this belongs in Thousand Token Wood

The app literally is a thousand-token wood. Every thought a creature thinks, every line it speaks, every overblown headline is generated under a strict token budget — and the wood counts them honestly in the corner of your screen. Remove the model and nothing remains: the AI isn't a feature, it's the climate.

How it works

Browser (hand-drawn SVG world, custom frontend) ── @gradio/client
        │ streamed events (thoughts, speech, moves, headlines)
gr.Server (FastAPI + Gradio queue)
        │
World engine (Python): places · items · weather · clock · consequences
        │ observe → think → act (ONE tool per turn)
llama-server (llama.cpp, pinned build) — fine-tuned Qwen3-1.7B Q4_K_M
        └─ per-creature KV slots + grammar-constrained JSON tool calls
  • —Agents, honestly small. Each creature turn is one grammar-constrained generation: {think, tool, to, arg}. llama.cpp compiles a JSON schema to a GBNF grammar, so a 1.7B model files tidy paperwork every single time — no retries, no parsing surgery.
  • —Fast enough to feel alive on 2 vCPUs. Static persona cards live in per-creature KV-cache slots; only the fresh world-slice is re-processed each turn. Median turn ≈ a few seconds, streamed to the UI as it happens.
  • —Emergence, not scripts. The newspaper is generated from the day's real event log. The misquotes are the model's own journalistic instincts.

Bonus quests claimed

BadgeEvidence
🔌 Off the GridAll inference is in-Space llama.cpp on CPU; zero external AI APIs
🎯 Well-TunedRuns a published fine-tune (in-voice agent SFT) — link above
🎨 Off-Brand100% custom frontend over gr.Server (SVG forest, zero default Gradio UI)
🦙 Llama Championllama.cpp llama-server, pinned release, grammar-constrained decoding
📡 Sharing is CaringFull agent traces (messages → decision → applied effects) published as a dataset; in-app 📡 button exports your own session
📓 Field NotesBlog post on building a living world on a 1.7B CPU budget — link above

Credits

Built for the Build Small Hackathon 2026 (Gradio × Hugging Face). Base model: Qwen3-1.7B (Apache-2.0). Runtime: llama.cpp. Fine-tuned with Modal GPU credits (see field notes for the training recipe). Every creature opinion is its own; the Tribune regrets nothing.