mdeberta-v3-base-name-classifier

This model is a fine-tuned version of microsoft/mdeberta-v3-base on ele-sage/person-company-names-classification dataset.

It achieves the following results on the evaluation set:

• Loss: 0.0305
• Accuracy: 0.9922
• Precision: 0.9957
• Recall: 0.9906
• F1: 0.9931

Model description

This model is a high-performance binary text classifier, fine-tuned from mdeberta-v3-base. Its purpose is to distinguish between a person's name and a company/organization name with high accuracy.

Direct Use

This model is intended to be used for text classification. Given a string, it will return a label indicating whether the string is a Person or a Company.

from transformers import pipeline

classifier = pipeline("text-classification", model="ele-sage/distilbert-base-uncased-name-classifier")

results = classifier([
    "Satya Nadella",
    "Global Innovations Inc.",
    "Martinez, Alonso"
])

for result in results:
    print(f"Text: '{result['text']}', Prediction: {result['label']}, Score: {result['score']:.4f}")

Downstream Use

This model is a key component of a two-stage name processing pipeline. It is designed to be used as a fast, efficient "gatekeeper" to first identify person names before passing them to a more complex parsing model, such as ele-sage/distilbert-base-uncased-name-splitter.

Out-of-Scope Use

• This model is not a general-purpose classifier. It is highly specialized for distinguishing persons from companies and will not perform well on other classification tasks (e.g., sentiment analysis).

Bias, Risks, and Limitations

• Geographic & Cultural Bias: The training data is heavily biased towards North American (Canadian) person names and Quebec-based company names. The model will be less accurate when classifying names from other cultural or geographic origins.
• Ambiguity: Certain names can legitimately be both a person's name and a company's name (e.g., "Ford"). In these cases, the model makes a statistical guess based on its training data, which may not always align with the specific context.
• Data Source: The person name data is derived from a Facebook data leak and contains noise. While a rigorous cleaning process was applied, the model may have learned from some spurious data.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-06
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Precision	Recall	F1
0.0592	0.0203	2000	0.0526	0.9877	0.9912	0.9872	0.9892
0.0473	0.0406	4000	0.0429	0.9891	0.9940	0.9868	0.9904
0.0491	0.0610	6000	0.0407	0.9893	0.9949	0.9863	0.9906
0.0383	0.0813	8000	0.0386	0.9898	0.9954	0.9868	0.9911
0.0415	0.1016	10000	0.0378	0.9904	0.9950	0.9881	0.9915
0.0315	0.1219	12000	0.0410	0.9905	0.9955	0.9877	0.9916
0.0416	0.1422	14000	0.0387	0.9908	0.9950	0.9888	0.9919
0.0292	0.1625	16000	0.0383	0.9908	0.9964	0.9874	0.9919
0.0381	0.1829	18000	0.0357	0.9907	0.9959	0.9878	0.9918
0.0266	0.2032	20000	0.0395	0.9909	0.9938	0.9902	0.9920
0.035	0.2235	22000	0.0392	0.9909	0.9956	0.9885	0.9920
0.0333	0.2438	24000	0.0356	0.9910	0.9935	0.9907	0.9921
0.0321	0.2641	26000	0.0343	0.9909	0.9947	0.9894	0.9920
0.0308	0.2845	28000	0.0360	0.9912	0.9954	0.9892	0.9923
0.0317	0.3048	30000	0.0348	0.9912	0.9941	0.9905	0.9923
0.0359	0.3251	32000	0.0346	0.9913	0.9959	0.9889	0.9924
0.0437	0.3454	34000	0.0333	0.9912	0.9957	0.9889	0.9923
0.0401	0.3657	36000	0.0334	0.9914	0.9954	0.9895	0.9924
0.0419	0.3861	38000	0.0321	0.9915	0.9957	0.9895	0.9926
0.032	0.4064	40000	0.0339	0.9914	0.9947	0.9902	0.9925
0.0367	0.4267	42000	0.0314	0.9916	0.9948	0.9904	0.9926
0.0276	0.4470	44000	0.0355	0.9915	0.9954	0.9897	0.9925
0.0373	0.4673	46000	0.0321	0.9916	0.9954	0.9899	0.9926
0.0364	0.4876	48000	0.0327	0.9915	0.9966	0.9885	0.9925
0.0317	0.5080	50000	0.0311	0.9914	0.9934	0.9915	0.9924
0.0355	0.5283	52000	0.0307	0.9917	0.9957	0.9898	0.9927
0.0276	0.5486	54000	0.0321	0.9918	0.9952	0.9904	0.9928
0.0342	0.5689	56000	0.0319	0.9918	0.9956	0.9900	0.9928
0.0316	0.5892	58000	0.0314	0.9918	0.9949	0.9906	0.9928
0.0322	0.6096	60000	0.0315	0.9916	0.9942	0.9912	0.9927
0.0357	0.6299	62000	0.0309	0.9921	0.9955	0.9905	0.9930
0.0296	0.6502	64000	0.0326	0.9919	0.9955	0.9903	0.9929
0.0324	0.6705	66000	0.0312	0.9919	0.9958	0.9900	0.9929
0.0266	0.6908	68000	0.0319	0.9920	0.9958	0.9902	0.9930
0.028	0.7112	70000	0.0321	0.9920	0.9961	0.9899	0.9930
0.0276	0.7315	72000	0.0319	0.9919	0.9963	0.9895	0.9929
0.0288	0.7518	74000	0.0316	0.9920	0.9952	0.9908	0.9930
0.0295	0.7721	76000	0.0304	0.9920	0.9955	0.9904	0.9930
0.0305	0.7924	78000	0.0309	0.9920	0.9963	0.9896	0.9929
0.0298	0.8127	80000	0.0312	0.9921	0.9962	0.9899	0.9930
0.0241	0.8331	82000	0.0312	0.9921	0.9954	0.9907	0.9930
0.0332	0.8534	84000	0.0308	0.9920	0.9955	0.9906	0.9930
0.0281	0.8737	86000	0.0301	0.9922	0.9957	0.9905	0.9931
0.0274	0.8940	88000	0.0305	0.9921	0.9952	0.9908	0.9930
0.0263	0.9143	90000	0.0300	0.9922	0.9958	0.9905	0.9931
0.0215	0.9347	92000	0.0304	0.9921	0.9952	0.9909	0.9931
0.0367	0.9550	94000	0.0297	0.9922	0.9956	0.9907	0.9931
0.0298	0.9753	96000	0.0302	0.9922	0.9955	0.9908	0.9931
0.0202	0.9956	98000	0.0305	0.9922	0.9957	0.9906	0.9931

Framework versions

• Transformers 4.57.1
• Pytorch 2.9.0+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1