tinyemo / README.md
anudit's picture
Release tinyemo model and inference code
c394b65 verified
|
Raw History Blame Contribute Delete
4.87 kB
metadata
language:
  - en
library_name: pytorch
tags:
  - emoji
  - text-classification
  - onnx
  - safetensors
base_model: desert-ant-labs/emo
license: other
license_name: desert-ant-labs-source-available
license_link: https://license.desertant.com/

tinyemo

Tinyemo is an English emoji suggestion model distilled from Emo. It predicts one of 812 emoji labels from a short phrase. It normalizes text to ASCII letters and digits and uses the last four words; it is intended for short input, not general sentiment analysis.

The evaluated distribution is tiny_emo.bin and tiny_tokenizer.bin, together 495,502 bytes. model.onnx and model.safetensors contain the same quantized-then-dequantized weights in float32, so they are larger; they are not int4/int8 files. They output logits, not text or emoji. The model uses custom preprocessing and is not a Transformers AutoModel or AutoTokenizer.

Usage

Download the files into a directory named tinyemo, then run Python from its parent directory. The reference runtime requires NumPy:

from tinyemo.runtime import TinyEmoRuntime

model = TinyEmoRuntime("tinyemo/tiny_emo.bin")
print(model.suggest("walk the dog", 3))

suggest returns (emoji, probability) pairs, defaults to one result, clamps the requested count to 1–3, and returns an empty list for input that normalizes to empty. Probabilities are softmax scores, not calibrated confidence estimates.

ONNX Runtime

Install numpy and onnxruntime. The supplied runtime provides the exact tokenizer, normalization and hashed n-gram preprocessing. config.json contains the label order and architecture settings.

import numpy as np
import onnxruntime as ort
from tinyemo.runtime import TinyEmoRuntime

pre = TinyEmoRuntime("tinyemo/tiny_emo.bin")
session = ort.InferenceSession("tinyemo/model.onnx", providers=["CPUExecutionProvider"])
text = "walk the dog"
ids, mask, buckets, signs = pre.encode_batch([text])
count = (signs[:, :, 0] != 0).sum(axis=1).astype(np.float32)
logits = session.run(None, {
    "tok_ids": ids, "tok_mask": mask, "ng_bucket": buckets,
    "ng_sign": signs.astype(np.float32), "ng_count": count,
})[0]
print(pre.labels[int(logits[0].argmax())])

The ONNX graph takes a dynamic batch dimension, with token arrays [batch, 16], n-gram arrays [batch, 96, 2], and ng_count [batch]. tok_ids and ng_bucket are int64, tok_mask is bool, and ng_sign and ng_count are float32. Preprocessing is outside the graph.

PyTorch and safetensors

Install torch, safetensors, and numpy:

import json
from safetensors.torch import load_file
from tinyemo.model import TinyEmo, TinyEmoConfig

with open("tinyemo/config.json", encoding="utf-8") as f:
    config = json.load(f)
model = TinyEmo(TinyEmoConfig(**config["config"]))
model.load_state_dict(load_file("tinyemo/model.safetensors"))
model.eval()

Use the same five preprocessed inputs shown above, converted to PyTorch tensors, to obtain logits. The Python implementation is included in this repository. export_release.py rebuilds both formats from the TEM2 file; it requires PyTorch and ONNX.

Evaluation and limitations

The original quantized runtime scored 87.7% top-1 (95% bootstrap interval 84.5–90.9%) and 97.1% top-3 (95.2–98.7%) on 375 hand-labelled phrases with accepted alternatives. Emo scored 88.8% and 97.1%, respectively. The paired top-1 difference was −1.1 percentage points (95% interval −4.3 to +2.1); this interval does not establish equivalence. Model selection used held-out training-pool agreement, while gold was logged during training.

On 6,000 held-out snippets from GoEmotions, DailyDialog and tweet_eval, top-1 agreement with the teacher was 62.2%; that measures agreement, not accuracy. Rare nouns such as “koala” remain a weakness. The gold set is small and author-labelled; these results may not generalize to other inputs. Earlier v1 results used a different test set.

Training data included short snippets from GoEmotions, DailyDialog, tweet_eval, SST-2, emotion, Tatoeba and Unicode CLDR, plus contexts built from the teacher's vocabulary. The gold phrases were excluded from the training pool, although constituent words could appear elsewhere. The model was trained on teacher predictions; no LLM was used for teacher labelling.

License and attribution

Tinyemo was distilled from Emo, which is distributed under the Desert Ant Labs Source-Available License. Review those terms and the licenses of the source datasets before distribution or commercial use. The other metadata records the teacher's license and does not assert additional rights for this derivative. See THIRD_PARTY_NOTICES.md for upstream notices. Training data is not included in this release folder.