text2emoji-tiny: 2.4M T5, q8 ONNX (Transformers.js layout)

README.md

@@ -0,0 +1,40 @@
+---
+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.

config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "T5ForConditionalGeneration"
+  ],
+  "classifier_dropout": 0.0,
+  "d_ff": 512,
+  "d_kv": 32,
+  "d_model": 128,
+  "decoder_start_token_id": 0,
+  "dense_act_fn": "relu",
+  "dropout_rate": 0.1,
+  "dtype": "float32",
+  "eos_token_id": 1,
+  "feed_forward_proj": "relu",
+  "initializer_factor": 1.0,
+  "is_decoder": false,
+  "is_encoder_decoder": true,
+  "is_gated_act": false,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "t5",
+  "num_decoder_layers": 3,
+  "num_heads": 4,
+  "num_layers": 3,
+  "pad_token_id": 0,
+  "relative_attention_max_distance": 128,
+  "relative_attention_num_buckets": 32,
+  "scale_decoder_outputs": true,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.17.0",
+  "use_cache": true,
+  "vocab_size": 8192
+}

generation_config.json

@@ -0,0 +1,8 @@
+{
+  "decoder_start_token_id": 0,
+  "eos_token_id": 1,
+  "pad_token_id": 0,
+  "max_new_tokens": 20,
+  "num_beams": 1,
+  "no_repeat_ngram_size": 1
+}

onnx/decoder_model_merged_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da45ee0e72df7e4d15a2a8ade7f394884708ca764f4c7eb86daefc42670d4429
+size 2128556

onnx/encoder_model_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2866c5d73b9878dff3d3a13b617821b7170c27c4b9454a46aafd25caabe9249
+size 1734642

special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "pad_token": "<pad>",
+  "eos_token": "</s>",
+  "unk_token": "<unk>"
+}

tokenizer_config.json

@@ -0,0 +1,11 @@
+{
+  "backend": "tokenizers",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 128,
+  "pad_token": "<pad>",
+  "tokenizer_class": "T5Tokenizer",
+  "unk_token": "<unk>"
+}