tokenaii
/

Neo

@@ -1,133 +1,83 @@
 ---
 license: other
-license_name: tokenai-neo-model-license-v3.0
-license_link: https://huggingface.co/tokenaii/Neo/blob/main/licenses/MODEL_LICENSE.md
-library_name: pytorch
-pipeline_tag: text-classification
 tags:
 - system-one
 - decision-model
 - rlcd
 - tool-routing
-- en
 ---
-# NEO - Decision Model
-<p align="center"><img src="assets/neo-cover.png" alt="NEO - Decision Model" width="360"></p>
-## Ownership and contact
-TokenAI is a non-profit startup founded in 2025 by Assem Sabry, based in
-Alexandria, Egypt. The Neo model, source code, training data, documentation,
-and related materials are owned by TokenAI.
-Contact: **info@tokenai.llc** · https://tokenai.llc
-## Summary
-Neo is a compact TokenAI System One decision model for selecting actions from a
-declared candidate set. It reads English application context and a typed
-decision schema, then returns calibrated outputs for `choice`, `score`, and
-`noul` in one encoder forward pass. It is not a conversational or free-form
-text-generation model.
-Canonical model: `tokenaii/Neo` — https://huggingface.co/tokenaii/Neo
-Training dataset: `tokenaii/Neo-dataset` — https://huggingface.co/datasets/tokenaii/Neo-dataset
-## Model architecture
-- Bidirectional Transformer encoder, initialized and trained from scratch.
-- Approximately 110 million parameters.
-- 8 Transformer layers; hidden size 512; 8 attention heads.
-- Intermediate size 2048; maximum context 1024 tokens.
-- Typed heads: `choice`, `score`, and `noul`.
-- Choice head: up to 32 candidate options per question.
-- Score head: up to 5 ordered levels.
-- Noul head: binary probability for a yes/no or escalation decision.
-- A request can contain up to 8 independent decision questions.
-## Input and output contract
-The input is English text plus a predefined decision schema and candidate
-options. The output contains typed probabilities and the selected index or
-level. Applications should apply their own confidence thresholds, abstention
-rules, validation, logging, and human review. Neo does not execute tools and
-does not replace authorization or policy enforcement.
-## Intended uses
-Non-commercial research and education for tool routing, workflow selection,
-request classification, department routing, escalation detection, validation
-gates, and selecting the next action in an agent pipeline.
-## Out-of-scope uses
-The license prohibits commercial, monetized, paid, sponsored, client-facing,
-production-business, or financially beneficial use. Do not use Neo for
-unsupervised medical, legal, financial, employment, housing, admissions,
-insurance, credit, safety-critical, government-benefit, law-enforcement, or
-irreversible decisions.
-## Training data and procedure
-The English-only corpus contains 400,000 synthetic records and 1,000,000
-independent decisions: 799,882 training, 100,138 calibration, and 99,980 test
-decisions. The catalog contains 20 tools; records commonly present one target
-and five distractors.
-The training run uses RLCD-inspired weighted soft-target training from scratch,
-20 epochs, approximately 15,640 optimizer steps, batch size 1024, bfloat16,
-learning rate `0.0002`, choice loss weight `8.0`, and score/noul weights `1.0`.
-The generator is identified as `neo-english-synthetic-v1` in the project
-manifests. The final benchmark report is produced after training and is not
-claimed by this card until verified.
-## Tokenizer
-Neo uses the standard English `bert-base-uncased` tokenizer and vocabulary. A
-custom tokenizer was not built for this release, so no custom tokenizer is
-claimed or published under `tokenizer/`. The Transformer parameters are Neo's
-own randomly initialized and trained weights; the tokenizer vocabulary is not
-the model architecture.
-## Release status
-The previous model weights were removed from this repository before the final
-English-only training release. The current repository contains documentation,
-configuration, tokenizer metadata, and licenses while the English training and
-automatic evaluation pipeline complete. Do not infer benchmark accuracy from
-the architecture or data counts. Verified results will be published only after
-the local evaluation artifacts are reviewed.
-## Repository layout
-- `licenses/MODEL_LICENSE.md` — full model license.
-- `licenses/NOTICE.md` — attribution and ownership notice.
-- `assets/neo-cover.png` — repository artwork.
-- `docs/MODEL_SPECIFICATION.md` — technical specification.
-- `tokenizer/` — reserved for a custom tokenizer only if one is actually built.
-- `training_config.json` — reproducibility metadata when a release includes it.
-## License and attribution
-Use is governed by [the TokenAI Neo Model License v3.0](licenses/MODEL_LICENSE.md).
-Redistribution, renaming, rebranding, white-labeling, Derivative Models,
-Commercial Use, and Financial Benefit are prohibited without written permission
-from TokenAI. Every permitted downstream report or model must state:
-> This work was trained using the TokenAI Neo Decision Model Dataset:
-> `tokenaii/Neo-dataset`.
-Permission requests must be sent to **info@tokenai.llc**. The complete legal
-terms, trademark policy, notice-and-takedown process, patent reservation,
-contribution/CLA rules, and dependency obligations are in the linked license.
-## Limitations
-Neo is trained on synthetic English decision records and may fail on unseen
-schemas, ambiguous contexts, distribution shifts, adversarial options, or
-languages other than English. Calibration and accuracy are task-dependent.
-Always validate with held-out, task-specific data and add human oversight for
-high-impact workflows.

 ---
 license: other
+license_name: tokenai-neo-dataset-license
+license_link: https://huggingface.co/datasets/tokenaii/Neo-dataset/blob/main/legal/DATASET_LICENSE.md
+language:
+- en
+task_categories:
+- text-classification
 tags:
 - system-one
 - decision-model
 - rlcd
+- synthetic
 - tool-routing
 ---
+# NEO - Decision Model Dataset
+<p align="center"><img src="assets/neo-cover.png" alt="NEO - Decision Model Dataset" width="360"></p>
+TokenAI is a non-profit startup founded in 2025 by Assem Sabry, based in Alexandria, Egypt. The dataset,
+model, source code, documentation, and related materials are owned by TokenAI.
+Contact: info@tokenai.llc · https://tokenai.llc
+Canonical reference: `tokenaii/Neo` — https://huggingface.co/tokenaii/Neo
+Complete source repository: [github.com/tokenaii/Neo](https://github.com/tokenaii/Neo).
+It contains the complete project source code, training code, synthetic-data
+generation code, evaluation and benchmark scripts, configurations, examples,
+documentation, licenses, and reproducibility materials. This repository
+contains the dataset files, manifest, provenance information, and dataset legal
+notices.
+Synthetic, schema-driven decision records for training and evaluating Neo, a System One decision model.
+## Contents
+- 1,000,000 individual decisions.
+- 400,000 state records.
+- `choice`, `score`, and `noul` questions.
+- English-only state text, questions, labels, and tool descriptions.
+- Tool routing, urgency, human-review, ambiguity, distractor tools, option permutations, and unknowable cases.
+## Provenance
+The records are generated by the `neo-synthetic` generator in the Neo repository. No rows from public datasets are copied into the generated corpus. Public datasets are used only for schema design, sanity checks, or separately documented evaluation.
+Generator: `neo-english-synthetic`.
+Generator seed: `20261003`.
+The complete source file used by training has SHA-256
+`3eff0bcf1ced0e748698fefefcf8069d2cc997888d7568fddb5ccd6a1a651c50`.
+The four files under `data/` are a byte-preserving concatenation of that
+training source.
+## Splits
+- `train`: 799,882 decisions
+- `calibration`: 100,138 decisions
+- `test`: 99,980 decisions
+The calibration split is reserved for probability calibration. The test split must not be used to select checkpoints, thresholds, or temperature.
+## Intended use and limitations
+This dataset is a research corpus for typed decision models. It is not a substitute for domain-specific validation, human review, or safety testing. Synthetic distributions can contain template artifacts and must not be treated as real user behavior.
+## Identity and license
+This dataset must remain identified as the **TokenAI Neo Decision Model Dataset** and must
+retain the canonical reference `tokenaii/Neo`. Rebranding, renaming,
+white-labeling, redistributing, rehosting, or publishing it under another
+identity is prohibited. The dataset must not be used without preserving its
+reference and provenance. See [the dataset license](legal/DATASET_LICENSE.md)
+and [the ownership notice](legal/NOTICE.md).
+Commercial, monetized, paid, sponsored, advertising-supported, production,
+client, resale, or financially beneficial use is forbidden. Any model or
+experiment trained with this dataset must clearly state that it used
+`tokenaii/Neo-dataset`.
+Third-party datasets referenced by the project are not included in this release
+and retain their own terms.