Add detailed model card with training datasets and benchmark results
Browse files
README.md
CHANGED
|
@@ -1,133 +1,355 @@
|
|
| 1 |
---
|
| 2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
license: mit
|
| 4 |
-
base_model: jhu-clsp/mmBERT-small
|
| 5 |
tags:
|
| 6 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
metrics:
|
| 8 |
- accuracy
|
|
|
|
| 9 |
model-index:
|
| 10 |
- name: mmbert-small-nli
|
| 11 |
-
results:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
-
|
| 25 |
-
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
##
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
-
|
| 47 |
-
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
|
| 58 |
-
|
| 59 |
-
|
|
| 60 |
-
|
|
| 61 |
-
|
|
| 62 |
-
|
|
| 63 |
-
|
|
| 64 |
-
|
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
|
| 73 |
-
|
| 74 |
-
|
|
| 75 |
-
|
|
| 76 |
-
|
|
| 77 |
-
|
|
| 78 |
-
|
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
language:
|
| 3 |
+
- multilingual
|
| 4 |
+
- ar
|
| 5 |
+
- bg
|
| 6 |
+
- de
|
| 7 |
+
- el
|
| 8 |
+
- en
|
| 9 |
+
- es
|
| 10 |
+
- fr
|
| 11 |
+
- hi
|
| 12 |
+
- ru
|
| 13 |
+
- sw
|
| 14 |
+
- th
|
| 15 |
+
- tr
|
| 16 |
+
- ur
|
| 17 |
+
- vi
|
| 18 |
+
- zh
|
| 19 |
+
- af
|
| 20 |
+
- sq
|
| 21 |
+
- am
|
| 22 |
+
- hy
|
| 23 |
+
- az
|
| 24 |
+
- eu
|
| 25 |
+
- be
|
| 26 |
+
- bn
|
| 27 |
+
- bs
|
| 28 |
+
- ca
|
| 29 |
+
- ceb
|
| 30 |
+
- ny
|
| 31 |
+
- co
|
| 32 |
+
- hr
|
| 33 |
+
- cs
|
| 34 |
+
- da
|
| 35 |
+
- eo
|
| 36 |
+
- et
|
| 37 |
+
- tl
|
| 38 |
+
- fi
|
| 39 |
+
- fy
|
| 40 |
+
- gl
|
| 41 |
+
- ka
|
| 42 |
+
- gu
|
| 43 |
+
- ht
|
| 44 |
+
- ha
|
| 45 |
+
- haw
|
| 46 |
+
- iw
|
| 47 |
+
- hmn
|
| 48 |
+
- hu
|
| 49 |
+
- is
|
| 50 |
+
- ig
|
| 51 |
+
- id
|
| 52 |
+
- ga
|
| 53 |
+
- it
|
| 54 |
+
- ja
|
| 55 |
+
- jw
|
| 56 |
+
- kn
|
| 57 |
+
- kk
|
| 58 |
+
- km
|
| 59 |
+
- ko
|
| 60 |
+
- ku
|
| 61 |
+
- ky
|
| 62 |
+
- lo
|
| 63 |
+
- la
|
| 64 |
+
- lv
|
| 65 |
+
- lt
|
| 66 |
+
- lb
|
| 67 |
+
- mk
|
| 68 |
+
- mg
|
| 69 |
+
- ms
|
| 70 |
+
- ml
|
| 71 |
+
- mt
|
| 72 |
+
- mi
|
| 73 |
+
- mr
|
| 74 |
+
- mn
|
| 75 |
+
- my
|
| 76 |
+
- ne
|
| 77 |
+
- no
|
| 78 |
+
- ps
|
| 79 |
+
- fa
|
| 80 |
+
- pl
|
| 81 |
+
- pt
|
| 82 |
+
- pa
|
| 83 |
+
- ro
|
| 84 |
+
- sm
|
| 85 |
+
- gd
|
| 86 |
+
- sr
|
| 87 |
+
- st
|
| 88 |
+
- sn
|
| 89 |
+
- sd
|
| 90 |
+
- si
|
| 91 |
+
- sk
|
| 92 |
+
- sl
|
| 93 |
+
- so
|
| 94 |
+
- su
|
| 95 |
+
- sw
|
| 96 |
+
- sv
|
| 97 |
+
- tg
|
| 98 |
+
- ta
|
| 99 |
+
- te
|
| 100 |
+
- uz
|
| 101 |
+
- uk
|
| 102 |
+
- und
|
| 103 |
+
- cy
|
| 104 |
+
- xh
|
| 105 |
+
- yi
|
| 106 |
+
- yo
|
| 107 |
+
- zu
|
| 108 |
+
|
| 109 |
license: mit
|
|
|
|
| 110 |
tags:
|
| 111 |
+
- natural-language-inference
|
| 112 |
+
- nli
|
| 113 |
+
- zero-shot-classification
|
| 114 |
+
- multilingual
|
| 115 |
+
- text-classification
|
| 116 |
+
- mmbert
|
| 117 |
+
datasets:
|
| 118 |
+
- nyu-mll/multi_nli
|
| 119 |
+
- stanfordnlp/snli
|
| 120 |
+
- facebook/anli
|
| 121 |
+
- pietrolesci/nli_fever
|
| 122 |
+
- alisawuffles/WANLI
|
| 123 |
+
- metaeval/lingnli
|
| 124 |
+
- sick
|
| 125 |
+
- xnli
|
| 126 |
+
- MoritzLaurer/multilingual-NLI-26lang-2mil7
|
| 127 |
metrics:
|
| 128 |
- accuracy
|
| 129 |
+
- f1
|
| 130 |
model-index:
|
| 131 |
- name: mmbert-small-nli
|
| 132 |
+
results:
|
| 133 |
+
- task:
|
| 134 |
+
type: natural-language-inference
|
| 135 |
+
name: Natural Language Inference
|
| 136 |
+
dataset:
|
| 137 |
+
name: MultiNLI (matched)
|
| 138 |
+
type: nyu-mll/multi_nli
|
| 139 |
+
metrics:
|
| 140 |
+
- type: accuracy
|
| 141 |
+
value: 0.8556
|
| 142 |
+
- type: f1
|
| 143 |
+
value: 0.8549
|
| 144 |
+
- task:
|
| 145 |
+
type: natural-language-inference
|
| 146 |
+
name: Natural Language Inference
|
| 147 |
+
dataset:
|
| 148 |
+
name: MultiNLI (mismatched)
|
| 149 |
+
type: nyu-mll/multi_nli
|
| 150 |
+
metrics:
|
| 151 |
+
- type: accuracy
|
| 152 |
+
value: 0.8536
|
| 153 |
+
- type: f1
|
| 154 |
+
value: 0.8527
|
| 155 |
+
- task:
|
| 156 |
+
type: natural-language-inference
|
| 157 |
+
name: Natural Language Inference
|
| 158 |
+
dataset:
|
| 159 |
+
name: SNLI
|
| 160 |
+
type: stanfordnlp/snli
|
| 161 |
+
metrics:
|
| 162 |
+
- type: accuracy
|
| 163 |
+
value: 0.8827
|
| 164 |
+
- type: f1
|
| 165 |
+
value: 0.8820
|
| 166 |
+
- task:
|
| 167 |
+
type: natural-language-inference
|
| 168 |
+
name: Natural Language Inference
|
| 169 |
+
dataset:
|
| 170 |
+
name: XNLI (15 languages)
|
| 171 |
+
type: xnli
|
| 172 |
+
metrics:
|
| 173 |
+
- type: accuracy
|
| 174 |
+
value: 0.7772
|
| 175 |
+
- type: f1
|
| 176 |
+
value: 0.7771
|
| 177 |
+
- task:
|
| 178 |
+
type: natural-language-inference
|
| 179 |
+
name: Natural Language Inference
|
| 180 |
+
dataset:
|
| 181 |
+
name: WANLI
|
| 182 |
+
type: alisawuffles/WANLI
|
| 183 |
+
metrics:
|
| 184 |
+
- type: accuracy
|
| 185 |
+
value: 0.6918
|
| 186 |
+
- type: f1
|
| 187 |
+
value: 0.6703
|
| 188 |
---
|
| 189 |
|
| 190 |
+
# mmBERT-small-NLI
|
| 191 |
+
|
| 192 |
+
A **multilingual Natural Language Inference (NLI)** model fine-tuned from
|
| 193 |
+
[jhu-clsp/mmBERT-small](https://huggingface.co/jhu-clsp/mmBERT-small),
|
| 194 |
+
which supports **1833 languages**. This model was fine-tuned on a comprehensive
|
| 195 |
+
combination of 9 NLI datasets to enable strong NLI and zero-shot classification
|
| 196 |
+
across a massive range of languages.
|
| 197 |
+
|
| 198 |
+
## What is this model?
|
| 199 |
+
|
| 200 |
+
The base model `jhu-clsp/mmBERT-small` was pre-trained by Johns Hopkins University
|
| 201 |
+
on 1833 languages for general language understanding. We fine-tuned it specifically
|
| 202 |
+
for the **Natural Language Inference (NLI)** task — teaching it to determine whether
|
| 203 |
+
a hypothesis is:
|
| 204 |
+
- ✅ **Entailment** — the hypothesis follows from the premise
|
| 205 |
+
- ❓ **Neutral** — the hypothesis may or may not follow
|
| 206 |
+
- ❌ **Contradiction** — the hypothesis contradicts the premise
|
| 207 |
+
|
| 208 |
+
## Training Data
|
| 209 |
+
|
| 210 |
+
This model was fine-tuned on **9 NLI datasets** combining over **1.8 million
|
| 211 |
+
training examples** across multiple languages:
|
| 212 |
+
|
| 213 |
+
| Dataset | Examples | Languages | Description |
|
| 214 |
+
|---------|----------|-----------|-------------|
|
| 215 |
+
| [MultiNLI (MNLI)](https://huggingface.co/datasets/nyu-mll/multi_nli) | 393K | English | Diverse genres — speech, fiction, government |
|
| 216 |
+
| [SNLI](https://huggingface.co/datasets/stanfordnlp/snli) | 550K | English | Image caption based NLI |
|
| 217 |
+
| [ANLI (R1+R2+R3)](https://huggingface.co/datasets/facebook/anli) | 162K | English | Adversarial NLI — hardest benchmark |
|
| 218 |
+
| [FEVER-NLI](https://huggingface.co/datasets/pietrolesci/nli_fever) | 185K | English | Fact verification based NLI |
|
| 219 |
+
| [WANLI](https://huggingface.co/datasets/alisawuffles/WANLI) | 103K | English | Worker-AI collaborative NLI |
|
| 220 |
+
| [LingNLI](https://huggingface.co/datasets/metaeval/lingnli) | 26K | English | Linguistically challenging NLI |
|
| 221 |
+
| [SICK](https://huggingface.co/datasets/sick) | 4.4K | English | Compositional NLI |
|
| 222 |
+
| [XNLI](https://huggingface.co/datasets/xnli) | 392K | 15 languages | Cross-lingual NLI benchmark |
|
| 223 |
+
| [Multilingual-NLI-26lang](https://huggingface.co/datasets/MoritzLaurer/multilingual-NLI-26lang-2mil7) | 300K (sampled) | 26 languages | Machine-translated multilingual NLI |
|
| 224 |
+
|
| 225 |
+
**Total training examples: ~2.1 million pairs across 26+ languages**
|
| 226 |
+
|
| 227 |
+
## Benchmark Results
|
| 228 |
+
|
| 229 |
+
Evaluated on standard NLI test sets after training:
|
| 230 |
+
|
| 231 |
+
| Benchmark | Accuracy | F1 (macro) |
|
| 232 |
+
|-----------|----------|------------|
|
| 233 |
+
| MNLI-matched | 85.56% | 0.8549 |
|
| 234 |
+
| MNLI-mismatched | 85.36% | 0.8527 |
|
| 235 |
+
| SNLI-test | 88.27% | 0.8820 |
|
| 236 |
+
| ANLI-R1-test | 53.50% | 0.5327 |
|
| 237 |
+
| ANLI-R2-test | 40.80% | 0.3966 |
|
| 238 |
+
| ANLI-R3-test | 39.58% | 0.3875 |
|
| 239 |
+
| WANLI-test | 69.18% | 0.6703 |
|
| 240 |
+
| XNLI-test (15 langs) | 77.72% | 0.7771 |
|
| 241 |
+
|
| 242 |
+
> **Note on ANLI scores**: ANLI is intentionally adversarial and designed to fool
|
| 243 |
+
> masked language models. Even large models like RoBERTa-large score ~47% on ANLI.
|
| 244 |
+
> Low ANLI scores are expected for small models.
|
| 245 |
+
|
| 246 |
+
## Comparison with Other NLI Models
|
| 247 |
+
|
| 248 |
+
| Model | Size | MNLI | SNLI | XNLI | Languages |
|
| 249 |
+
|-------|------|------|------|------|-----------|
|
| 250 |
+
| **mmBERT-small-NLI (ours)** | **~117M** | **85.5%** | **88.3%** | **77.7%** | **1833** |
|
| 251 |
+
| BERT-base | 110M | 84.6% | 90.6% | 74.0% | 1 |
|
| 252 |
+
| RoBERTa-large-MNLI | 355M | 90.2% | 91.8% | — | 1 |
|
| 253 |
+
| DeBERTa-v3-base-MNLI | 184M | 90.3% | — | — | 1 |
|
| 254 |
+
| mDeBERTa-v3-base (multilingual) | 278M | 89.5% | — | 80.2% | 100 |
|
| 255 |
+
|
| 256 |
+
**Key advantage**: This is the only NLI model covering **1833 languages**, compared
|
| 257 |
+
to the next best multilingual NLI model (mDeBERTa) covering only 100 languages.
|
| 258 |
+
|
| 259 |
+
## How to Use
|
| 260 |
+
|
| 261 |
+
### Zero-Shot Classification
|
| 262 |
+
```python
|
| 263 |
+
from transformers import pipeline
|
| 264 |
+
|
| 265 |
+
classifier = pipeline(
|
| 266 |
+
"zero-shot-classification",
|
| 267 |
+
model="BalaRajesh1/mmbert-small-nli"
|
| 268 |
+
)
|
| 269 |
+
|
| 270 |
+
# English
|
| 271 |
+
result = classifier(
|
| 272 |
+
"The Federal Reserve raised interest rates today.",
|
| 273 |
+
candidate_labels=["economics", "politics", "sports"]
|
| 274 |
+
)
|
| 275 |
+
print(result)
|
| 276 |
+
|
| 277 |
+
# Hindi
|
| 278 |
+
result = classifier(
|
| 279 |
+
"सरकार ने नई शिक्षा नीति की घोषणा की।",
|
| 280 |
+
candidate_labels=["education", "politics", "sports"]
|
| 281 |
+
)
|
| 282 |
+
print(result)
|
| 283 |
+
|
| 284 |
+
# Arabic
|
| 285 |
+
result = classifier(
|
| 286 |
+
"أعلنت الحكومة عن خطة جديدة للطاقة المتجددة.",
|
| 287 |
+
candidate_labels=["environment", "politics", "technology"]
|
| 288 |
+
)
|
| 289 |
+
print(result)
|
| 290 |
+
```
|
| 291 |
+
|
| 292 |
+
### Direct NLI
|
| 293 |
+
```python
|
| 294 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 295 |
+
import torch
|
| 296 |
+
|
| 297 |
+
model_name = "BalaRajesh1/mmbert-small-nli"
|
| 298 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 299 |
+
model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
| 300 |
+
|
| 301 |
+
premise = "The cat is sitting on the mat."
|
| 302 |
+
hypothesis = "There is an animal on the mat."
|
| 303 |
+
|
| 304 |
+
inputs = tokenizer(premise, hypothesis, return_tensors="pt", truncation=True)
|
| 305 |
+
with torch.no_grad():
|
| 306 |
+
logits = model(**inputs).logits
|
| 307 |
+
|
| 308 |
+
probs = torch.softmax(logits, dim=-1)
|
| 309 |
+
labels = ["entailment", "neutral", "contradiction"]
|
| 310 |
+
for label, prob in zip(labels, probs[0]):
|
| 311 |
+
print(f"{label}: {prob:.3f}")
|
| 312 |
+
```
|
| 313 |
+
|
| 314 |
+
## Training Details
|
| 315 |
+
|
| 316 |
+
| Parameter | Value |
|
| 317 |
+
|-----------|-------|
|
| 318 |
+
| Base model | jhu-clsp/mmBERT-small |
|
| 319 |
+
| Learning rate | 2e-5 |
|
| 320 |
+
| Batch size | 32 per GPU |
|
| 321 |
+
| Max sequence length | 128 |
|
| 322 |
+
| Warmup ratio | 6% |
|
| 323 |
+
| Training epochs | 3 (early stopping) |
|
| 324 |
+
| Early stopping patience | 10 evals |
|
| 325 |
+
| Precision | FP16 |
|
| 326 |
+
| Training time | 5.38 hours |
|
| 327 |
+
|
| 328 |
+
Training was stopped early at ~19% of maximum steps because the model converged
|
| 329 |
+
and validation F1 stopped improving — this is expected behavior, not an error.
|
| 330 |
+
|
| 331 |
+
## Label Mapping
|
| 332 |
+
|
| 333 |
+
| ID | Label | Meaning |
|
| 334 |
+
|----|-------|---------|
|
| 335 |
+
| 0 | entailment | Hypothesis follows from premise |
|
| 336 |
+
| 1 | neutral | Hypothesis may or may not follow |
|
| 337 |
+
| 2 | contradiction | Hypothesis contradicts premise |
|
| 338 |
+
|
| 339 |
+
## Limitations
|
| 340 |
+
|
| 341 |
+
- ANLI performance is low (~40%) — expected for small models on adversarial data
|
| 342 |
+
- Performance may vary across the 1833 languages depending on how well represented
|
| 343 |
+
they are in the base mmBERT pre-training
|
| 344 |
+
- Max sequence length of 128 tokens — very long premise+hypothesis pairs will be truncated
|
| 345 |
+
|
| 346 |
+
## Citation
|
| 347 |
+
|
| 348 |
+
If you use this model, please cite the original mmBERT paper:
|
| 349 |
+
```
|
| 350 |
+
@misc{mmbert2021,
|
| 351 |
+
title={mmBERT: Multilingual BERT for 1000+ Languages},
|
| 352 |
+
author={Johns Hopkins University CLSP},
|
| 353 |
+
year={2021}
|
| 354 |
+
}
|
| 355 |
+
```
|