| license: mit | |
| library_name: transformers.js | |
| pipeline_tag: text2text-generation | |
| language: | |
| - en | |
| tags: | |
| - emoji | |
| - t5 | |
| - onnx | |
| datasets: | |
| - KomeijiForce/Text2Emoji | |
| # text2emoji-tiny | |
| A 2.4M-parameter T5 trained from scratch by [runonweb](https://runonweb.ai/models/emoji) to turn an English | |
| sentence into a short emoji sequence. 3.9 MB as q8 ONNX (encoder 1.7 MB + merged decoder 2.1 MB), in the | |
| Transformers.js layout. | |
| | Config | Params | val loss | emoji-set F1 | | |
| |--------|--------|----------|--------------| | |
| | 3+3 layers, d_model 128, d_ff 512, 4 heads, BPE vocab 8192 | 2.43M | 3.72 | 0.38 | | |
| Data: [KomeijiForce/Text2Emoji](https://huggingface.co/datasets/KomeijiForce/Text2Emoji), ~493k | |
| ChatGPT-generated sentence → emoji pairs after cleaning. The dataset card lists no license; the weights are | |
| MIT. Recipe: `training/text2emoji` in the runonweb repo. | |
| ## Use | |
| ```ts | |
| import { Emojifier } from 'runonweb/emoji' | |
| const emojifier = new Emojifier() | |
| await emojifier.load() | |
| const { text } = await emojifier.emojify('I love pizza and my dog') // "🍕❤️🐶" | |
| ``` | |
| Or with Transformers.js directly: `pipeline('text2text-generation', 'midudev/text2emoji-tiny', { dtype: 'q8' })`, | |
| decoding with `no_repeat_ngram_size: 1`; emojis come back space-separated. | |