docs: add Neo Ar cards and licenses
#1
by assemsabry - opened
README.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: tokenai-neo-ar-model-license
|
| 4 |
+
license_link: https://huggingface.co/tokenaii/Neo-Ar/blob/main/licenses/MODEL_LICENSE.md
|
| 5 |
+
library_name: pytorch
|
| 6 |
+
pipeline_tag: text-classification
|
| 7 |
+
language:
|
| 8 |
+
- ar
|
| 9 |
+
tags:
|
| 10 |
+
- system-one
|
| 11 |
+
- decision-model
|
| 12 |
+
- rlcd
|
| 13 |
+
- tool-routing
|
| 14 |
+
- arabic
|
| 15 |
+
- arabic-only
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# Neo Ar — Arabic Decision Model
|
| 19 |
+
|
| 20 |
+
<p align="center"><img src="assets/neo-ar-cover.png" alt="Neo Ar — Arabic Decision Model" width="360"></p>
|
| 21 |
+
|
| 22 |
+
Neo Ar is an Arabic-only TokenAI System One decision model. It selects bounded
|
| 23 |
+
actions from declared candidate sets and returns typed probabilities for
|
| 24 |
+
`choice`, `score`, and `noul`. It does not generate free-form conversation.
|
| 25 |
+
|
| 26 |
+
## Language policy
|
| 27 |
+
|
| 28 |
+
The model weights, tokenizer, training records, labels, questions, tool
|
| 29 |
+
descriptions, and evaluation cases are Arabic-only. This model card and the
|
| 30 |
+
repository documentation are English for accessibility; English is not a
|
| 31 |
+
supported model language.
|
| 32 |
+
|
| 33 |
+
## Architecture target
|
| 34 |
+
|
| 35 |
+
- 12-layer bidirectional Transformer encoder.
|
| 36 |
+
- Hidden size 768; 12 attention heads.
|
| 37 |
+
- Intermediate size 3,072; maximum context 1,024 tokens.
|
| 38 |
+
- Up to 64 choice options and 7 ordered score levels.
|
| 39 |
+
- Binary `noul` head and up to 12 questions per request.
|
| 40 |
+
- 86,393,160 trainable parameters with the released 644-token vocabulary.
|
| 41 |
+
|
| 42 |
+
## Links
|
| 43 |
+
|
| 44 |
+
- Dataset: https://huggingface.co/datasets/tokenaii/Neo-Ar-Dataset
|
| 45 |
+
- Complete source and training repository: https://github.com/tokenaii/Neo-Ar
|
| 46 |
+
|
| 47 |
+
The GitHub repository contains the complete source code, Arabic data generator,
|
| 48 |
+
tokenizer construction, training code, evaluation, benchmark results, and
|
| 49 |
+
training history. This model repository contains release artifacts and public
|
| 50 |
+
documentation only.
|
| 51 |
+
|
| 52 |
+
## Status and limitations
|
| 53 |
+
|
| 54 |
+
This is a research model until held-out Arabic benchmarks demonstrate reliable
|
| 55 |
+
tool routing, calibration, option-order robustness, and unseen-tool behavior.
|
| 56 |
+
Do not use it for unsupervised high-impact or irreversible decisions.
|
| 57 |
+
|
| 58 |
+
TokenAI is a non-profit startup founded in 2025 by Assem Sabry, based in
|
| 59 |
+
Alexandria, Egypt. Contact: info@tokenai.llc — https://tokenai.llc.
|