# Thousand Token Wood

*Submission for the Build Small Hackathon, Thousand Token Wood track. June 2026.*

Five woodland creatures run a living market. Each one thinks on a different lab's
small model, and nobody scripts the drama: they trade, gossip, hoard, form grudges,
and panic on their own. You play the Patron, a shadow financier who profits from the
chaos. And the central discovery of the build is that the wood fights back.

- **Play it:** https://huggingface.co/spaces/build-small-hackathon/thousand-token-wood-sim
- **Demo video (100s):** https://youtu.be/ugFyFumUCgs
- **Social post:** https://x.com/RealLesterLeong/status/2064904967391961483

## The council: five minds, four labs, 29.5B parameters

The honest version of "small models can run an economy" is not one model wearing five
hats. It is five distinct minds with five distinct temperaments arguing a market into
being:

| Creature | Model | Lab |
|---|---|---|
| Oona the owl | gpt-oss-20b | OpenAI |
| Fenn the fox | Nemotron-Mini-4B | NVIDIA |
| Bramble the squirrel | MiniCPM3-4B | OpenBMB |
| Mossback the tortoise and Pip the mouse | ttw-trader-0.5b | fine-tuned (ours) |

*The council. Distinct-engine budget 29.5B, under the 32B cap. Every thought and raw
JSON completion is exposed in the UI, so you can check that none of it is scripted.*

The 0.5B is a LoRA fine-tune I distilled from a 3B teacher with the mistakes stripped
out of the training set. The student trades cleaner than its teacher: zero self-buys
and 100% valid offers, against the teacher's 2.2% self-buy rate. Small, done
carefully, beat bigger.

## The game: insider trading as a decaying resource

You are not a benevolent god. You short a good, whisper a tip to set up its fall,
spring a Wood Legend (a famous market panic reskinned as woodland folklore), and
collect when the price craters. A live exposure meter shows your expected payoff
before you commit.

The first time, it works: the crash pays 55 pebbles, exactly as authored. But the
creatures you burn remember. They sour on you, hoard against your next crash, and
testify to Magistrate Heron. Run the identical gambit again and the meter forecasts
16 pebbles, and it pays exactly 16. Then the verdicts land: the owl is exiled, your
pebbles are frozen. A trick you repeat is a trick that dies.

That mechanic is the build's thesis about agents, not just its plot: emergent behavior
from one model population evaporates when you change the cast, you cannot steer a
heterogeneous council by shocking its inputs, and a repeated manipulation gets priced
in by the agents on the other side. The full journey, including the two failed builds
that taught me the most, is in the field notes below.

## Why small is load-bearing

A living economy needs many agents thinking many times per turn, plus a narrator.
Frontier models are too slow and too costly for that loop. Four small engines, each on
its own GPU with per-engine batching and scale-to-zero, make a real-time council
feasible: the entire project, including fine-tuning, four engine deployments, trace
publication, and three recorded story arcs, used about a quarter of the starter Modal
credits.

## Built with

- **Gradio** Space (custom storybook UI, cinematic title card, instant attract-mode
  replay so the Space is never blank while engines wake)
- **Modal** end to end: four vLLM engines as separate apps, LoRA training, evals
- **Models:** gpt-oss-20b, Nemotron-Mini-4B, MiniCPM3-4B, and a published 0.5B
  fine-tune ([AdmiralTaco/ttw-trader-0.5b](https://huggingface.co/AdmiralTaco/ttw-trader-0.5b))

## Bonus quests

- **Well-Tuned:** the app serves the published fine-tune above, with before/after
  reliability evals in the field notes.
- **Sharing is Caring:** [open agent traces](https://huggingface.co/datasets/build-small-hackathon/thousand-token-wood-traces),
  every row tagged with the lab and model that produced it, so you can compare how
  four labs' small models read the same market state.
- **Off-Brand:** custom illustrated town square, lab-colored attribution, operator
  console chrome, and a pure-CSS cinematic open.
- **Field Notes,** in four parts:
  1. [Emergent market drama](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim)
  2. [The five-lab council](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v2)
  3. [The crash that vanished](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v3)
  4. [The wood fights back](https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v4)

I build market models professionally, and this little wood taught me real lessons
about agent economies at a scale where the only thing at risk was a pile of pebbles.
If it makes you smile, a like on the Space is the community vote.

Small models, big adventures.