johnpaulbin/toxicity-setfit-5-large

SetFit with jhu-clsp/mmBERT-base

This is a SetFit model that can be used for Text Classification. This SetFit model uses jhu-clsp/mmBERT-base as the Sentence Transformer embedding model. A LogisticRegression instance is used for classification.

The model has been trained using an efficient few-shot learning technique that involves:

1. Fine-tuning a Sentence Transformer with contrastive learning.
2. Training a classification head with features from the fine-tuned Sentence Transformer.

Model Details

Model Description

• Model Type: SetFit
• Sentence Transformer body: jhu-clsp/mmBERT-base
• Classification head: a LogisticRegression instance
• Maximum Sequence Length: 8192 tokens
• Number of Classes: 2 classes

Model Sources

• Repository: SetFit on GitHub
• Paper: Efficient Few-Shot Learning Without Prompts
• Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts

Model Labels

Label	Examples
toxic	'slanteye' 'lets meet in real life cutie ;)' 'youre dead to me and soon irl'
not toxic	'wanna make a snowman?' 'word' 'boost'

Evaluation

Metrics

Label	Accuracy
all	0.9970

Uses

Direct Use for Inference

First install the SetFit library:

pip install setfit

Then you can load this model and run inference.

from setfit import SetFitModel

# Download from the 🤗 Hub
model = SetFitModel.from_pretrained("johnpaulbin/toxicity-setfit-5-large")
# Run inference
preds = model("cooked him")

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	1	4.6067	81

Label	Training Sample Count
not toxic	8689
toxic	4512

Training Hyperparameters

• batch_size: (128, 128)
• num_epochs: (1, 1)
• max_steps: -1
• sampling_strategy: num_iterations
• num_iterations: 12
• body_learning_rate: (2e-05, 1e-05)
• head_learning_rate: 0.01
• loss: CosineSimilarityLoss
• distance_metric: cosine_distance
• margin: 0.25
• end_to_end: False
• use_amp: True
• warmup_proportion: 0.1
• l2_weight: 0.01
• seed: 42
• eval_max_steps: -1
• load_best_model_at_end: True

Training Results

Epoch	Step	Training Loss	Validation Loss
0.0004	1	0.3044	-
0.0202	50	0.2803	-
0.0404	100	0.2355	-
0.0606	150	0.1607	-
0.0808	200	0.0911	-
0.1010	250	0.0694	-
0.1212	300	0.0532	-
0.1414	350	0.0391	-
0.1616	400	0.0343	-
0.1817	450	0.0296	-
0.2019	500	0.0264	-
0.2221	550	0.022	-
0.2423	600	0.0201	-
0.2625	650	0.0127	-
0.2827	700	0.0165	-
0.3029	750	0.0097	-
0.3231	800	0.0093	-
0.3433	850	0.0098	-
0.3635	900	0.009	-
0.3837	950	0.0076	-
0.4039	1000	0.0073	-
0.4241	1050	0.0076	-
0.4443	1100	0.0058	-
0.4645	1150	0.0049	-
0.4847	1200	0.0066	-
0.5048	1250	0.0044	-
0.5250	1300	0.0052	-
0.5452	1350	0.0041	-
0.5654	1400	0.0061	-
0.5856	1450	0.0033	-
0.6058	1500	0.0051	-
0.6260	1550	0.0029	-
0.6462	1600	0.0039	-
0.6664	1650	0.0046	-
0.6866	1700	0.0028	-
0.7068	1750	0.0037	-
0.7270	1800	0.0022	-
0.7472	1850	0.0032	-
0.7674	1900	0.0025	-
0.7876	1950	0.0027	-
0.8078	2000	0.0024	-
0.8279	2050	0.0031	-
0.8481	2100	0.0024	-
0.8683	2150	0.0033	-
0.8885	2200	0.0039	-
0.9087	2250	0.0028	-
0.9289	2300	0.0029	-
0.9491	2350	0.0023	-
0.9693	2400	0.0022	-
0.9895	2450	0.0027	-
1.0	2476	-	0.0046

Framework Versions

• Python: 3.12.12
• SetFit: 1.2.0.dev0
• Sentence Transformers: 5.2.0
• Transformers: 4.57.3
• PyTorch: 2.9.0+cu126
• Datasets: 4.0.0
• Tokenizers: 0.22.2

Citation

BibTeX

@article{https://doi.org/10.48550/arxiv.2209.11055,
    doi = {10.48550/ARXIV.2209.11055},
    url = {https://arxiv.org/abs/2209.11055},
    author = {Tunstall, Lewis and Reimers, Nils and Jo, Unso Eun Seo and Bates, Luke and Korat, Daniel and Wasserblat, Moshe and Pereg, Oren},
    keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {Efficient Few-Shot Learning Without Prompts},
    publisher = {arXiv},
    year = {2022},
    copyright = {Creative Commons Attribution 4.0 International}
}