OpenAI vs GLiNER PII Comparison

openai/privacy-filter and GLiNER large-v2.1 were compared on 600 PII samples (400 English, 200 multilingual).

• openai/privacy-filter: 1.5B total parameters, sparse MoE with only 50M activated per forward pass
• GLiNER large-v2.1: a zero-shot model with about 300M parameters
• CPU throughput: openai/privacy-filter 2.8 samples/sec, GLiNER 1.1 samples/sec

On English macro F1 under strict exact match, GLiNER scored 0.367 versus openai/privacy-filter's 0.155, but under boundary overlap, openai/privacy-filter led with 0.498 versus GLiNER's 0.416. Most of the gap came from 1-character offset differences caused by the GPT-style BPE tokenizer, making it more of a span alignment issue than actual misses.

By category under the boundary criterion, openai/privacy-filter outperformed on PERSON, EMAIL, PHONE, DATE, while GLiNER led on ADDRESS. In particular, EMAIL was effectively solved, with 0.987 for English and 1.000 for multilingual.

GLiNER's F1 was about 8 points higher at a threshold of 0.7 compared to the default 0.5. In conclusion, GLiNER is favorable when minimizing misses is important, while openai/privacy-filter is favorable when precision and CPU throughput matter, and GLiNER's zero-shot input approach is better suited when custom entity types are needed.

However, openai/privacy-filter requires trust_remote_code=True and a development branch of transformers, making it difficult to use directly with a stable release yet.