Nirnaya attribution notice

Nirnaya-Flash code and weights are released under the Apache License, Version 2.0 (see LICENSE).

The weights derive from the following model:

1. ModernBERT (Apache-2.0)
   Source: answerdotai/ModernBERT-large
   Revision: 45bb4654a4d5aaff24dd11d4781fa46d39bf8c13
   Use: encoder architecture and pretrained encoder weights.

The weights were trained on the following datasets. Each remains under its own license.

2. Typed Decisions (Apache-2.0)
   Source: LocalLLaMA/typed-decisions @ c76749ec58bd8c3d2ea706b31c333a9059c38f90
   Attribution: Typed Decisions by LocalLLaMA.

3. Open-Jev (CC0-1.0)
   Source: ZefanCai/Open-Jev @ c67699e13d0ae25e35b77165a4b6b079bedc8aba
   Attribution: Open-Jev typed decision datasets by Zefan Cai (customer-control-v1 excluded).

4. BANKING77 (CC-BY-4.0)
   Source: PolyAI/banking77 @ 90d4e2ee5521c04fc1488f065b8b083658768c57
   Attribution: BANKING77 by PolyAI; Casanueva et al. (2020), "Efficient Intent Detection with
   Dual Sentence Encoders".

5. BoolQ (CC-BY-SA-3.0)
   Source: google/boolq @ 35b264d03638db9f4ce671b711558bf7ff0f80d5
   Attribution: BoolQ by Google Research; Clark et al. (2019), "BoolQ: Exploring the Surprising
   Difficulty of Natural Yes/No Questions". The dataset itself is not redistributed here.

6. UltraFeedback (MIT)
   Source: openbmb/UltraFeedback @ 40b436560ca83a8dba36114c22ab3c66e43f6d5e
   Attribution: UltraFeedback by OpenBMB; Cui et al. (2023).

7. GoEmotions (Apache-2.0)
   Source: google-research-datasets/go_emotions @ add492243ff905527e67aeb8b80c082af02207c3
   Attribution: GoEmotions by Demszky et al. (ACL 2020), Google Research.

8. News Category Dataset (CC-BY-4.0)
   Source: heegyu/news-category-dataset @ 304a05a55bc6abc0446d8fae0d0771716b6a271a
   Attribution: News Category Dataset by Rishabh Misra (2022), arXiv:2209.11429; HuffPost
   headlines and short descriptions.

9. Prompt Injections (Apache-2.0)
   Source: deepset/prompt-injections @ 4f61ecb038e9c3fb77e21034b22511b523772cdd
   Attribution: prompt-injections dataset by deepset.

Evaluation-only data (AG News, DAIR Emotion, SST-5, and the test splits above) was used to
measure the model and was never used to train the released weights.

This notice does not replace the original licenses of the works listed above.
