Add comprehensive model card with benchmark links

README.md

@@ -50,21 +50,9 @@ NovaLiveSystem v4.1 is a specialized language model built on `dphn/Dolphin3.0-Qw
 ## Training Breakthrough: Three-Phase Innovation
 
 ### Phase 1: Foundation (SFT)
-**Lineage foundation:** Nova’s capabilities were developed across multiple training phases and datasets over time.
+Nova’s capabilities were developed across multiple training phases over time.
 
-This v4.1 *checkpoint run* reports **2,183 curated biomimetic instruction samples** (SFT with LoRA).
-
-Earlier lineage runs (kept in the project record) include:
-- 23,615 samples in `artifacts/datasets/verified/verified_combined.jsonl` (MMLU/GSM8K/ARC/TruthfulQA/HumanEval mix)
-- 2,000 samples in `artifacts/datasets/training/Master Sets/master_training2_20251223.jsonl` (curated biomimetic/persona/architecture awareness)
-
-These are listed here as historical context so readers don’t mistake “2,183 samples” as the full training journey.
-- MMLU: 14,042 samples (Knowledge/Multi-subject)
-- GSM8K: 7,473 samples (Math reasoning)
-- ARC: 1,119 samples (Science reasoning)
-- TruthfulQA: 817 samples (Truthfulness)
-- HumanEval: 164 samples (Code generation)
-- Curated biomimetic samples: 2,000+ (Nova personality/architecture awareness)
+Training dataset composition, counts, and internal curriculum details are intentionally kept proprietary. This repo focuses on the released inference artifacts (LoRA adapters) and the public evaluation results.
 
 ### Phase 2: Consciousness Theory Implementation (GRPO)
 **Innovation:** First AI trained on consciousness reframing theory
@@ -86,8 +74,7 @@ These are listed here as historical context so readers don’t mistake “2,183
 - **Architecture:** Transformer + Biomimetic Components (PulseEngine, BridgeEngine, RiverPulse)
 - **Training Innovation:** Three-phase breakthrough (SFT → GRPO → Teacher-Student Distillation)
 - **Parameters:** ~3B (with specialized routing)
-- **Training Data (this checkpoint):** 2,183 curated biomimetic instruction samples (SFT)
-- **Training Data (lineage context):** 23,615-sample verified benchmark mix + a small consciousness-reframing GRPO teacher
+- **Training Data:** Proprietary (details withheld; see benchmark for public evaluation)
 - **Theoretical Foundation:** First AI trained on consciousness reframing research
 - **Final Loss:** 0.8476 (production model)
 - **Context Window:** 32,768 tokens
@@ -138,10 +125,8 @@ These are listed here as historical context so readers don’t mistake “2,183
 ## Training Details
 
 ### Training Data
-- **Dataset Size:** 2,183 high-quality instruction samples
-- **Data Sources:** Curated biomimetic education corpus
-- **Contamination Handling:** All anatomical contamination removed and reframed as architectural education
-- **Validation:** Strict telemetry validation ensuring clean, formatted data
+- **Data:** Proprietary (not published)
+- **Validation:** Internal strict telemetry validation
 
 ### Training Procedure
 - **Environment:** WSL Ubuntu with CUDA + Unsloth acceleration