Update README.md

README.md

@@ -42,93 +42,3 @@ Input -> [Embed] -> [Block x 4] -> [RMSNorm] -> [LM Head] -> Output
 | Flash Attention 2 | Stanford | O(1) memory fused attention kernel |
 | torch.compile | PyTorch 2.0+ | Graph compilation with operator fusion |
 
-## Specs
-
-| Parameter | Value |
-|-----------|-------|
-| Layers | 4 |
-| Embed dim | 256 |
-| Heads | 4 |
-| MLA latent | 32 |
-| FFN hidden | 682 |
-| Context | 256 |
-| Sliding window | 128 |
-| Vocab | 50,257 (GPT-2 BPE) |
-| Params | 15,586,816 |
-| Checkpoint | ~31 MB (FP16) |
-
-## Files
-
-```
-train_goat_gpt_nano.py    Training script (auto-downloads dataset)
-app.py                     HF Spaces deployment (Gradio UI)
-requirements.txt           Dependencies
-goat_gpt_nano.ht          Trained model output (generated)
-```
-
-## Build / Run
-
-### Train
-
-```bash
-pip install -r requirements.txt
-python train_goat_gpt_nano.py
-```
-
-Outputs:
-- `goat_gpt_nano.ht` -- deployable model
-- `goat_gpt_nano_gen.pt` -- generation checkpoint
-- `best.pt` -- best validation checkpoint
-- `ckpt_*.pt` -- intermediate checkpoints
-
-### Deploy (Hugging Face Spaces)
-
-1. Create Space -> Gradio SDK
-2. Upload:
-   - `goat_gpt_nano.ht`
-   - `app.py`
-   - `requirements.txt`
-
-## Training Config
-
-| Parameter | Value |
-|-----------|-------|
-| Batch size | 16 |
-| Grad accum | 2 (effective = 32) |
-| Max steps | 5000 |
-| Warmup | 250 |
-| Stable | 3000 |
-| Decay | 1750 |
-| LR | 8e-4 -> 8e-5 |
-| Optimizer | Lion |
-| Precision | FP16 |
-| Compile | default (T4-safe) |
-
-## Dataset
-
-- Source: WikiText-2 (auto-downloaded from HuggingFace)
-- Fallback chain: wikitext -> wikitext-2-raw-v1 -> Salesforce/wikitext -> local file
-- Train tokens: ~2.4M
-- Val tokens: ~250K
-
-## GPU
-
-- Recommended: T4 (Google Colab)
-- Cost: ~5-10 compute units per run
-- Time: ~3-5 hours
-- VRAM: <2 GB
-
-## Expected Results
-
-| Metric | Value |
-|--------|-------|
-| Val loss | ~3.5-4.7 |
-| Val PPL | ~33-110 |
-| Throughput | ~50-80K tok/s |
-
-## Notes
-
-- `TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=0` set before torch import (T4 compatibility)
-- `torch.amp.GradScaler('cuda')` and `torch.amp.autocast('cuda')` used (PyTorch 2.6+)
-- `trust_remote_code` removed (datasets library deprecated)
-- Warning suppression active for clean logs