notice: attribution to RF-DETR + WebPII
Browse files
NOTICE
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
screenpipe-pii-redactor
|
| 2 |
+
Copyright 2026 Mediar, Inc.
|
| 3 |
+
|
| 4 |
+
This product is a derivative work of the OpenAI Privacy Filter
|
| 5 |
+
(https://github.com/openai/privacy-filter), licensed under the
|
| 6 |
+
Apache License, Version 2.0. The full upstream Apache 2.0 license
|
| 7 |
+
text is preserved in LICENSE.upstream-apache2.txt.
|
| 8 |
+
|
| 9 |
+
The base model architecture, tokenizer, and pretrained weights are the
|
| 10 |
+
work of OpenAI and are licensed Apache 2.0. screenpipe-pii-redactor
|
| 11 |
+
extends those weights via supervised fine-tuning on a custom corpus and
|
| 12 |
+
re-initializes the output head for a 12-label PII taxonomy specific to
|
| 13 |
+
desktop activity logs.
|
| 14 |
+
|
| 15 |
+
Significant modifications introduced by this derivative:
|
| 16 |
+
- Output head re-initialized for a 12-class PII label space
|
| 17 |
+
(29 rows copied from upstream where labels aligned, 20 rows
|
| 18 |
+
initialized from random for new classes; see model/finetune_summary.json).
|
| 19 |
+
- Fine-tuned for 3 epochs on a mixed corpus of:
|
| 20 |
+
* synthetic accessibility / window-title / OCR data
|
| 21 |
+
(private — not redistributed)
|
| 22 |
+
* a 25% slice of ai4privacy/pii-masking-300k (CC-BY-4.0)
|
| 23 |
+
with labels mapped to the 12-class taxonomy
|
| 24 |
+
* targeted secret-shape augmentation (private)
|
| 25 |
+
- Context window n_ctx raised from 128 to 256.
|
| 26 |
+
- Hyperparameters: batch_size=4, lr=1e-4, weight_decay=0,
|
| 27 |
+
max_grad_norm=1.0, shuffle_seed=1337.
|
| 28 |
+
|
| 29 |
+
Distribution license:
|
| 30 |
+
|
| 31 |
+
- The fine-tuned weights and accompanying materials in this repository
|
| 32 |
+
(the "Derivative Work") are licensed under CC BY-NC 4.0; see LICENSE.
|
| 33 |
+
- The Apache 2.0 obligations on the base model are preserved by:
|
| 34 |
+
(a) shipping LICENSE.upstream-apache2.txt with this repo,
|
| 35 |
+
(b) attributing OpenAI Privacy Filter in README.md,
|
| 36 |
+
(c) declaring significant modifications above.
|
| 37 |
+
- "OpenAI" and "Privacy Filter" are trademarks / brands of OpenAI; this
|
| 38 |
+
derivative does not use those marks to endorse or suggest endorsement
|
| 39 |
+
of this work by OpenAI.
|
| 40 |
+
|
| 41 |
+
Third-party datasets used during fine-tuning:
|
| 42 |
+
|
| 43 |
+
- ai4privacy/pii-masking-300k
|
| 44 |
+
https://huggingface.co/datasets/ai4privacy/pii-masking-300k
|
| 45 |
+
Licensed CC-BY-4.0.
|
| 46 |
+
|
| 47 |
+
Questions about license compatibility or commercial use:
|
| 48 |
+
louis@screenpi.pe
|