bilune/heptapod-ink

Heptapod Ink

Weights for Heptapod Ink, a hobby experiment inspired by the film Arrival. Two networks learned a ring-shaped writing system made of real ink marks cut from the film's logograms. A writer turns an English sentence into one logogram, and a reader recovers the sentence from the image alone.

Try it in the browser: Heptapod Ink Space.

Files

File	What it is
model.pt	Full PyTorch checkpoint (v4, stage s2): T5-small encoder, slot writer, codebooks, eyes, adapter, T5-small decoder. Load with v4/heptapod_t5.py from the code repo.
web/writer_alt_mix.onnx	Writer: tokens to the 16 nearest symbols per slot (int8 weights, fp16 vocabulary).
web/render.onnx	Canonical renderer: 36 mark ids to a 288 px ink image.
web/eyes_q8pc.onnx	Eyes: ink image to mark probabilities per slot.
web/memory.onnx	Adapter for reading images.
web/sym_memory.onnx	Adapter for scoring candidate symbols while rereading.
web/decoder_fp16.onnx	T5 decoder with tied vocabulary, used both for reading and for rereading.
web/config.json, web/tokenizer/	Ring config, symbol to mark table, T5 tokenizer.

How well it reads

Locked exam on 2,000 natural sentences from Universal Dependencies English treebanks, scored once:

Mode	Exact sentence	Words
Plain writing	33.5%
Writer rereads before sending	81.0%	92.9%
Rereading, noisy image	75.5%
Control: shuffled image	1.2%

Short sentences are easy (3 to 5 words: 94.4% exact) and long ones much harder (9 to 12 words: 67.9%). Random word salad mostly fails, and so do characters T5 does not know (for example "ñ").

Limits

English only, at most about 12 words per logogram. The code is learned for model-to-model use; humans are not expected to read it.

License and credits

CC BY-NC 4.0, because the ink comes from the Arrival logograms by Martine Bertrand, released by Wolfram Research under CC BY-NC 4.0. T5-small by Google (Apache 2.0). Training data: Tatoeba, COGS, FineWeb, Wikipedia, CC-News. See NOTICE.md in the code repo.