Upload NOTICE with huggingface_hub

NOTICE

@@ -1,16 +1,17 @@
 Omnex
 Copyright 2026 Shivam Parmar
 
-This product includes software and model weights developed by third parties.
-The notices below are required by the licenses of those components and must be
-retained in any redistribution, including redistribution of derived model
-weights.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use
+this model except in compliance with the License. You may obtain a copy of the
+License at:
+
+    http://www.apache.org/licenses/LICENSE-2.0
 
 --------------------------------------------------------------------------------
-BASE MODEL
+ATTRIBUTION
 --------------------------------------------------------------------------------
 
-Omnex is a derivative work of Qwen3-8B.
+Omnex is a derivative work of:
 
     Qwen3-8B
     Copyright 2025 Alibaba Cloud / Qwen Team
@@ -18,59 +19,11 @@ Omnex is a derivative work of Qwen3-8B.
     Licensed under the Apache License, Version 2.0
 
 CHANGES MADE (Apache-2.0 Section 4(b)):
-    Qwen3-8B was adapted by LoRA supervised fine-tuning (rank 16, alpha 32,
-    applied to q/k/v/o and gate/up/down projections), then merged back into the
-    base weights at 16-bit precision. Training used loss masking over assistant
-    responses only. No architectural change was made; the tokenizer and chat
-    template are unmodified. The adaptation specializes the model for
-    retrieval-grounded question answering with citations and abstention.
-
-You may obtain a copy of the Apache License 2.0 at:
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
---------------------------------------------------------------------------------
-TRAINING DATA
---------------------------------------------------------------------------------
-
-    SQuAD 2.0 (rajpurkar/squad_v2)
-    Rajpurkar, Jia and Liang, "Know What You Don't Know: Unanswerable Questions
-    for SQuAD", ACL 2018.
-    Licensed under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)
-
-    HotpotQA (hotpot_qa, distractor setting)
-    Yang, Qi, Zhang, Bengio, Cohen, Salakhutdinov and Manning, "HotpotQA: A
-    Dataset for Diverse, Explainable Multi-hop Question Answering", EMNLP 2018.
-    Licensed under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)
-
-Passages and answers from these datasets were reformatted into cited and
-abstaining response templates. Derived data inherits CC BY-SA 4.0 obligations:
-attribution must be preserved and derivative datasets shared under the same
-license.
-
-Additional training data (synthetic edge cases, advanced case types, and the
-identity set) was generated by this project and is covered by the project
-license.
-
---------------------------------------------------------------------------------
-TOOLING
---------------------------------------------------------------------------------
-
-    Unsloth — https://github.com/unslothai/unsloth — Apache License 2.0
-    TRL (Transformer Reinforcement Learning), Hugging Face — Apache License 2.0
-    Transformers, Datasets, Hugging Face Hub — Apache License 2.0
-    PyTorch — BSD-3-Clause
-    vLLM — Apache License 2.0
-    ChromaDB — Apache License 2.0
-
---------------------------------------------------------------------------------
-EVALUATION REFERENCES
---------------------------------------------------------------------------------
-
-Models evaluated for comparison are the property of their respective owners and
-are not redistributed here. They were used solely for benchmarking:
 
-    Phi-4 (Microsoft), Llama 3.1 8B Instruct (Meta), Claude Sonnet (Anthropic)
+    The base model was adapted by supervised fine-tuning to specialize it for
+    retrieval-grounded question answering with inline citations and abstention.
+    No architectural change was made; the tokenizer and chat template are
+    unmodified.
 
-Answer scoring follows the normalization and token-F1 method of the official
-SQuAD 2.0 evaluation script.
+Training data, training methodology and hyperparameters are proprietary and are
+not published.