9e25412dc0
The app forgives slips to a neighbouring key unless the person turns it off, so English's evidence is measured that way: on the 252 Turk Dialogues replies with what the partner said, 51.2% among the five sentences and 31.7% first (52.8% and 32.9% with slips not forgiven), above the 50 bar. The catalogue's credits link where each piece is published. Co-Authored-By: Claude Opus 5.5 <[email protected]>
3.1 KiB
3.1 KiB
First letters: licences
First letters in AuraSense AAC (aac-web) downloads one small language model, on request, and runs it on the person's device, on its graphics chip: it turns the first letter of each word they type into whole sentences for them to choose from. Nothing typed is sent anywhere.
Qwen3 0.6B, taught first letters (English)
- The model: Qwen3-0.6B-Base by the Qwen team (Alibaba Cloud), fine-tuned by AuraSense to turn first letters into sentences and stored in 4 bits for WebGPU (https://huggingface.co/Qwen/Qwen3-0.6B-Base). Licence: Apache-2.0 (https://www.apache.org/licenses/LICENSE-2.0); the text is
licences/Apache-2.0-Qwen3.832dd9e00a68.txt. - What it learned from: Turk AAC communications by Keith Vertanen and Per Ola Kristensson (“The Imagination of Crowds: Conversational AAC Language Modeling using Crowdsourcing and Large Data Sources”, EMNLP 2011), turned into first-letter examples (https://aactext.org/imagine/). Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
- What it learned from: Synthetic-Persona-Chat by Pegah Jandaghi, XiangHai Sheng, Xinyi Bai, Jay Pujara and Hakim Sidahmed (Google, 2023), its consecutive turns turned into first-letter examples (https://huggingface.co/datasets/google/Synthetic-Persona-Chat). Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
- What it learned from: Conversations written for it by AuraSense with Gemma 4 26B-A4B (Google), whose outputs are free to use (https://ai.google.dev/gemma). Licence: Apache-2.0 (https://www.apache.org/licenses/LICENSE-2.0).
- Source:
huggingface.co/Qwen/Qwen3-0.6B-Base@da87bfb608, fine-tuned by AuraSense to turn first letters into sentences (LoRA rank 16 on every linear layer, merged; 281,414 examples), exported to INT4 (block 16) for WebGPU with onnxruntime-genai 0.17.1, its graph edited so every node runs on WebGPU. - What was changed: the model's weights by the fine-tuning above; the training data's sentences were turned into examples (each reply's first letters, what the partner said before it, some with typing slips), and nothing of them is shipped. The graph was edited so every node runs on WebGPU in onnxruntime-web 1.30 (its attention-mask arithmetic fed in directly, the search's caches reordered on the graphics chip). 14 files of at most 64 MiB, each named by its own SHA-256 (393 MB).
- The way the letters are written (each word's first letter, contractions split at the apostrophe, punctuation where it was typed) follows S. Cai et al., “Context-Aware Abbreviation Expansion Using Large Language Models”, NAACL 2022 (the method only; none of its data is used).
- Tested on Turk Dialogues (K. Vertanen, 2017), which is used for testing only and is not shipped.
- Ready in: English (en).
The runtime that runs it (onnxruntime-web, MIT) is the listener's, vendored with its licence, notice and ONNX Runtime's third-party notices under vendor/onnxruntime-web-<version>-<revision>/; the tokenizer library (@huggingface/tokenizers, Apache-2.0) and the app's worker are vendored with their licence and notice under vendor/hf-tokenizers-<version>-<revision>/.