tde-general-v0.2

TDE general v0.2 — the typed-decision encoder, now also playing Snake

The same 150M encoder as tde-general-v0.1 (ModernBERT-base, pointer readout; one forward pass over state + question + candidates returns a probability per candidate), fine-tuned further so it plays Snake from a plain-text board while keeping its general decisions.

Snake

The state is the board as text, one token per cell ( . empty, F food, H head, 1– 9 body: moves until that cell is free) under a line with the length and heading; the question is "Which way should the snake move?" with the candidates up, down, left and right. The top candidate is the move. An anti-trap shield vetoes a move after which the head can no longer reach its tail and takes the model's next choice instead.

About 12 ms per move with MLX on an M5 Pro. A live page where the model plays game after game: python -m integrations.snake.demo in the code repository (tutorial: integrations/snake).

General decisions

JevBench and the typed-decisions test were not re-run for this checkpoint.

Training

Initialised from tde-general-v0.1 and fine-tuned for 2 epochs (6,250 steps of 32) on 100k Snake positions from 8x8, 12x12 and 24x16 boards, each labelled with the moves of a shortest-path planner that keeps its tail reachable. Trained with MLX on one Mac in about 80 minutes (integrations/snake/build_sft.py, then tde/mlx/train.py).

Usage

# pip install git+https://github.com/e13ven-arch/tde   (package name: tde)
from tde.inference import Decider
d = Decider.from_run("tdelab/tde-general-v0.2")   # or a local release folder
d.decide(board_text, {"type": "choice", "instructions": "Which way should the snake move?",
                      "criteria": {m: m for m in ("up", "down", "left", "right")}})

General questions work exactly as with v0.1.

Limitations

Without the shield the snake eventually traps itself: all four protocol games end before move 600. General accuracy is about one point below v0.1; the limitations listed for v0.1 apply unchanged.