docs: keep training records on GitHub only

README.md

@@ -1,6 +1,6 @@
 ---
 license: other
-license_name: tokenai-neo-model-license-v3.0
+license_name: tokenai-neo-model-license
 license_link: https://huggingface.co/tokenaii/Neo/blob/main/licenses/MODEL_LICENSE.md
 library_name: pytorch
 pipeline_tag: text-classification
@@ -36,6 +36,11 @@ Canonical model: `tokenaii/Neo` — https://huggingface.co/tokenaii/Neo
 
 Training dataset: `tokenaii/Neo-dataset` — https://huggingface.co/datasets/tokenaii/Neo-dataset
 
+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.
+
 ## Model architecture
 
 - Bidirectional Transformer encoder, initialized and trained from scratch.
@@ -70,36 +75,31 @@ unsupervised medical, legal, financial, employment, housing, admissions,
 insurance, credit, safety-critical, government-benefit, law-enforcement, or
 irreversible decisions.
 
-## Training data and procedure
+## Training records
+
+Training code, configurations, live logs, loss history, checkpoints metadata,
+evaluation reports, and benchmark history are maintained only in the GitHub
+source repository:
 
-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.
+https://github.com/tokenaii/Neo/tree/main/docs/training
 
-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.
+This Hugging Face model repository does not publish the training history or
+benchmark archive.
 
 ## 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.
+Neo uses the standard English `bert-base-uncased` tokenizer and vocabulary,
+published under `tokenizer/` for repository organization. A custom tokenizer
+was not built. 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.
+documentation, tokenizer metadata, and licenses. Do not infer benchmark
+accuracy from the architecture or data counts. Training and benchmark records
+are maintained only in the GitHub source repository.
 
 ## Repository layout
 
@@ -107,12 +107,11 @@ the local evaluation artifacts are reviewed.
 - `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.
+- `tokenizer/` — standard English `bert-base-uncased` tokenizer files.
 
 ## License and attribution
 
-Use is governed by [the TokenAI Neo Model License v3.0](licenses/MODEL_LICENSE.md).
+Use is governed by [the TokenAI Neo Model License](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: