fix: add quantization models and performance

README.md

@@ -15,6 +15,7 @@ language:
   - ja
   - tt
 license: openrail++
+library: transformers
 base_model:
   - distilbert/distilbert-base-multilingual-cased
 datasets:
@@ -34,6 +35,8 @@ We perform an 85/15 train-test split per language based on the `textdetox` datas
 
 ## Performance Overview
 
+### Original model
+
 | Language   | Validation F1 | Training F1 |   ΔF1   |
 |------------|---------------|-------------|---------|
 | Russian    | 0.9572        | 0.9897      | -0.0324 |
@@ -52,6 +55,47 @@ We perform an 85/15 train-test split per language based on the `textdetox` datas
 | Hinglish   | 0.7234        | 0.8533      | -0.1299 |
 | Hebrew     | 0.6455        | 0.8441      | -0.1987 |
 
+### Quantized model (ONNX)
+
+| Language  | Original Val F1 | Quantized Val F1 | Δ Val F1    |
+| --------- | --------------- | ---------------- | ----------- |
+| Russian   | 0.9572          | 0.9609           | **+0.0037** |
+| English   | 0.9528          | 0.9495           | **−0.0033** |
+| German    | 0.8904          | 0.8842           | **−0.0062** |
+| Hindi     | 0.9248          | 0.8940           | **−0.0300** |
+| French    | 0.9446          | 0.9351           | **−0.0095** |
+| Japanese  | 0.8658          | 0.8584           | **−0.0074** |
+| Ukrainian | 0.8997          | 0.8988           | **−0.0009** |
+| Tatar     | 0.9200          | 0.9148           | **−0.0052** |
+| Amharic   | 0.6513          | 0.6377           | **−0.0136** |
+| Spanish   | 0.8564          | 0.8439           | **−0.0125** |
+| Chinese   | 0.6865          | 0.6697           | **−0.0168** |
+| Arabic    | 0.7563          | 0.7535           | **−0.0028** |
+| Italian   | 0.8223          | 0.8033           | **−0.0190** |
+| Hinglish  | 0.7234          | 0.7260           | **+0.0026** |
+| Hebrew    | 0.6455          | 0.6190           | **−0.0265** |
+
+| Language  | Original Train F1 | Quantized Train F1 | Δ Train F1  |
+| --------- | ----------------- | ------------------ | ----------- |
+| Russian   | 0.9897            | 0.9875             | **−0.0022** |
+| English   | 0.9853            | 0.9857             | **+0.0004** |
+| German    | 0.9547            | 0.9369             | **−0.0178** |
+| Hindi     | 0.9599            | 0.9335             | **−0.0264** |
+| French    | 0.9874            | 0.9814             | **−0.0060** |
+| Japanese  | 0.9253            | 0.9081             | **−0.0172** |
+| Ukrainian | 0.9511            | 0.9476             | **−0.0035** |
+| Tatar     | 0.9682            | 0.9631             | **−0.0051** |
+| Amharic   | 0.6915            | 0.6863             | **−0.0052** |
+| Spanish   | 0.9399            | 0.9273             | **−0.0126** |
+| Chinese   | 0.7807            | 0.7596             | **−0.0211** |
+| Arabic    | 0.8550            | 0.8481             | **−0.0069** |
+| Italian   | 0.9271            | 0.9193             | **−0.0078** |
+| Hinglish  | 0.8533            | 0.8436             | **−0.0097** |
+| Hebrew    | 0.8441            | 0.8204             | **−0.0237** |
+
+ The quantized model maintains high accuracy and generalization on the evaluation set, 
+ with negligible performance loss for most languages.
+
 ## License
 
 This model is licensed under [OpenRAIL++](LICENSE)

model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bee01a9d2d474a589c5db631d99c382d5502bb81a0fe318d247f46c16b25f060
+size 541416974

model.quant.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cea601d7d45fa1e8fbc8c716127c9efff710339a00bea422d6f2565537b4fd80
+size 135898271