hsnsahinhsn commited on
Commit
8bbdb37
·
verified ·
1 Parent(s): a7ea41c
Files changed (1) hide show
  1. NOTICE +16 -15
NOTICE CHANGED
@@ -1,15 +1,16 @@
1
- Pusula — Turkish decision models
2
- Copyright 2026 Hasan Şahin
3
-
4
- This product includes or builds on the following Apache-2.0 licensed works:
5
-
6
- - d3 (Decision 3.0), https://huggingface.co/vllm-sr/d3
7
- Copyright 2026 vLLM Semantic Router Team. Built on Qwen/Qwen3.8-27B (Apache-2.0).
8
- The d3 prompt contract and non-causal attention hook in pusula/d3model.py follow vllm-sr/d3
9
- (adapted there from perplexity-ai/pplx-decider-v1.1-27b, Apache-2.0).
10
- - Qwen3.8-27B, https://huggingface.co/Qwen/Qwen3.8-27B (Qwen Team, Alibaba Cloud).
11
- - TabiBERT, https://huggingface.co/boun-tabilab/TabiBERT (Boğaziçi University TABILAB).
12
-
13
- Pusula model weights are LoRA adapters and a code readout trained on top of vllm-sr/d3 (Pusula),
14
- and a fine-tuned TabiBERT cross-encoder (Pusula Nano). Their training data was generated
15
- synthetically; it contains no scraped or personal data.
 
 
1
+ Pusula — Turkish decision models
2
+ Copyright 2026 Hasan Şahin
3
+
4
+ This product includes or builds on the following Apache-2.0 licensed works:
5
+
6
+ - d3 (Decision 3.0), https://huggingface.co/vllm-sr/d3
7
+ Copyright 2026 vLLM Semantic Router Team. Built on Qwen/Qwen3.8-27B (Apache-2.0).
8
+ The d3 prompt contract and non-causal attention hook in pusula/d3model.py follow vllm-sr/d3
9
+ (adapted there from perplexity-ai/pplx-decider-v1.1-27b, Apache-2.0).
10
+ - Qwen3.8-27B, https://huggingface.co/Qwen/Qwen3.8-27B (Qwen Team, Alibaba Cloud).
11
+ - Qwen3.5-0.8B, https://huggingface.co/Qwen/Qwen3.5-0.8B (Qwen Team, Alibaba Cloud).
12
+ - TabiBERT, https://huggingface.co/boun-tabilab/TabiBERT (Boğaziçi University TABILAB).
13
+
14
+ Pusula model weights are LoRA adapters and a code readout trained on top of vllm-sr/d3 (Pusula),
15
+ a fully fine-tuned Qwen3.5-0.8B in the d3 format (Pusula Mini) and a fine-tuned TabiBERT cross-encoder
16
+ (Pusula Nano). Their training data was generated synthetically; it contains no scraped or personal data.