carbonpoint/botc-artist

botc-automod Artist translator

A 135M-parameter model, fine-tuned from SmolLM2-135M-Instruct, that translates a player's yes/no question in Blood on the Clocktower into one query in a small JSON language. It is used by botc-automod when the automated storyteller answers the Artist.

The model does not answer questions. The game engine evaluates the query against the true game state, and the player confirms the engine's plain-English reading before the answer is given. A misread question can be rephrased; it cannot produce a false answer without the player seeing the reading first.

Input and output

System prompt: Translate the Blood on the Clocktower yes/no question into a JSON query.

User message:

Players (clockwise): Ann, Ben, Cat, Dan, Eve
Asker: Ann
Characters: chef=Chef, imp=Imp, ...
Question: is the person on my left evil?

Output: {"query":{"op":"is_team","player":"cw:me","team":"evil"}}, or {"query":{"op":"unanswerable"}} for questions that are not yes/no about the current game.

Training

LoRA (rank 64, all linear layers), 30,000 generated examples, 2 epochs, learning rate 2e-4. Examples pair random game worlds (3,288 real and invented player names, all three base editions) with templated questions and augmentation (typos, chatty openings, table talk). About 12% of questions are unanswerable.

Evaluation

Set	Right	Misread	Refused
Generated test (300; unseen wording and names)	90.0%	4.0%	6.0%
Hand-written (103; casual, typos, 20 unanswerable)	94.2%	3.9%	1.9%

8-bit GGUF (145 MB) with llama.cpp on a 4-core CPU: median 0.3 to 0.8 seconds per question. Some hand-written failure patterns informed later training templates (not the questions themselves), so the generated test set is the more independent measure. The 360M sibling scores 91.3% / 95.1% with slightly more misreads (6.7% / 3.9%) at 2.5 times the latency.

Files

• botc-artist.gguf: 8-bit (Q8_0), 145 MB
• SHA256SUMS: checksum