Update README.md

README.md

@@ -22,32 +22,32 @@ Running under the **JiRackTrain** pipeline on the enterprise infrastructure clus
 ## Model Details
 
 - **Model Name**: TernaryTransformer3B (3.3 Billion Parameters)
-- **Architecture**: TernaryTransformer (custom ternary bit-response logic with GQA + MoE scaling foundations)
-- **Tokenizer**: [JiRack Pro Tokenizer (128K)](https://huggingface.co/CMSManhattan/JiRack-Pro-Tokenizer-128K)
-  - 347 active language editions of Wikipedia
-  - Specialized robotic action tokens + traditional text/vision embeddings
+- **Architecture**: TernaryTransformer (custom ternary bit-response logic)
+- **Tokenizer**: [JiRack Pro Tokenizer (128K)](https://huggingface.co/CMSManhattan/JiRack-Pro-Tokenizer-128K) — 347 active language editions of Wikipedia + specialized robotic action tokens
 - **Pre-training Dataset**: [JiRack-Pretrain-Dataset](https://huggingface.co/datasets/CMSManhattan/JiRack-Pretrain-Dataset)
-- **Training Objective**: From-scratch pre-training using the exact same multimodal corpus used to optimize the JiRack Pro Tokenizer (perfect vocabulary alignment)
 
-## Why This Alignment Matters
+## Key Metrics (Step ~838/1000 on Shard 1)
 
-By reusing the tokenizer's own training corpus for model pre-training, JiRack Robotics eliminates standard vocabulary bias found in most open-source LLMs. The ternary neural paths learn directly from a data distribution that maps perfectly to the token vocabulary boundaries. This is expected to deliver significantly higher deterministic accuracy for real-time robotic control policies, edge processing, and low-latency physical maneuver loops.
+- **Iteration Speed**: ~242–352 seconds per iteration (network tensor loading)
+- **Current Loss**: 2.9678
+- **Moving Average Loss**: 6.0920
+- **Perplexity (PPL)**: 442.33
 
-## Training Setup & Technical Specifications
+Engineers anticipate a **steep drop** in the perplexity curve within the first 250 iterations of Shard 1 as the ternary weights align with the tokenizer's token distributions.
 
-The training leverages the custom `train_jirack_accelerate.py` framework with the following key paradigms:
+## Intended Use
 
-- **Memory Optimization**: Active **Gradient Checkpointing** across all TernaryTransformer3B layers (drastically reduced VRAM footprint)
-- **Data Layout**: Sequential pipeline distributed across **7 cloud shards** (`jirack_pretrain_chunk_0.pt` through `jirack_pretrain_chunk_6.pt`)
-- **Infrastructure**: Enterprise cluster with network-mounted storage (`/mnt/nfs_clientshare/JiRackTrain`)
-- **Framework**: PyTorch + Accelerate (Python 3.12)
+This foundational model is being developed specifically for **robotics applications**:
 
-**Current Training Log Snapshot** (Shard 1/7):
+- Real-time control policies
+- Multimodal reasoning (text + vision + action tokens)
+- Edge deployment with ternary efficiency
+- Low-latency physical interaction loops
 
-```bash
-Shards found for processing: 7
-Initializing JiRack architecture (3.3B) from your class...
--> Gradient Checkpointing successfully activated on TernaryTransformer3B.
-Starting sequential training loop across shards...
+## Monitoring & Updates
 
-[Shard 1/7] Loading over network: jirack_pretrain_chunk_0.pt
+Training is actively progressing across all **7 shards**. Follow this repository or the linked tokenizer/dataset cards for checkpoint releases, evaluation results, and fine-tuned robotic variants.
+
+---
+
+**Stay tuned** — the ternary robotics revolution is just getting started! 🚀