Upload folder using huggingface_hub

README.md

@@ -0,0 +1,122 @@
+# braille256-v5: Multimodal Universal Braille Model
+
+The first language model trained on **multimodal data encoded as 8-dot Braille**.
+
+## Key Innovation
+
+This model demonstrates that **any data type** (text, images, audio, binary) can be encoded into 8-dot Braille Unicode (U+2800-U+28FF) and processed by a single unified model.
+
+```python
+def byte_to_braille(byte: int) -> str:
+    """Direct 1:1 mapping: 256 bytes → 256 Braille patterns"""
+    return chr(0x2800 + byte)
+```
+
+## Model Details
+
+| Property | Value |
+|----------|-------|
+| Parameters | 11.5M |
+| Architecture | Transformer (4 layers, 4 heads) |
+| Hidden Size | 256 |
+| Vocabulary | 32,000 (SentencePiece Unigram) |
+| Context Length | 512 tokens |
+| Training Steps | 5,000 |
+| Final Loss | 2.87 |
+
+## Training Data
+
+- **Text**: 218 files from Project Gutenberg (7 languages) encoded as 8-dot Braille
+- **Audio**: Synthetic WAV files encoded as Braille
+- **Total**: 2M tokens from multimodal corpus
+
+## Compression
+
+| Tokenizer | Vocab | Chars/Token |
+|-----------|-------|-------------|
+| braille256-v4 (8k) | 8,192 | 2.24 |
+| **braille256-v5 (32k)** | 32,000 | **2.45** |
+| GPT-4 (reference) | 100,000 | 4.31 |
+
+## Universal Encoding
+
+The 8-dot Braille encoding enables:
+
+1. **Text** → UTF-8 bytes → Braille
+2. **Images** → Raw bytes → Braille  
+3. **Audio** → WAV bytes → Braille
+4. **Any file** → Bytes → Braille
+
+### Modality Headers
+
+```
+⣿⠁ = TEXT
+⣿⠃ = IMAGE
+⣿⠇ = AUDIO
+⣿⠏ = BINARY
+```
+
+## Usage
+
+```python
+import torch
+import sentencepiece as spm
+
+# Load tokenizer
+sp = spm.SentencePieceProcessor()
+sp.load("tokenizer.model")
+
+# Load model
+from train_multimodal_v5 import Braille256MultimodalModel, MultimodalConfig
+import json
+
+with open("config.json") as f:
+    config = MultimodalConfig.from_dict(json.load(f))
+
+model = Braille256MultimodalModel(config)
+model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
+model.eval()
+
+# Encode any data as Braille
+def bytes_to_braille(data: bytes) -> str:
+    return ''.join(chr(0x2800 + b) for b in data)
+
+# Generate from Braille prompt
+braille_text = "⠞⠓⠑⠀⠟⠥⠊⠉⠅"  # "the quick" in Braille
+tokens = sp.encode(braille_text)
+input_ids = torch.tensor([tokens])
+output = model.generate(input_ids, max_length=50)
+generated = sp.decode(output[0].tolist())
+print(generated)
+```
+
+## Model Family
+
+| Version | Type | Patterns | Focus |
+|---------|------|----------|-------|
+| v3 | 6-dot | 64 | Literary Braille |
+| v4 | 8-dot | 256 | Computer Braille |
+| **v5** | 8-dot | 256 | **Multimodal** |
+
+## Why 8-dot Braille?
+
+- **256 patterns** = exactly 1 byte
+- **Universal encoding**: Any data → Braille
+- **Tactile AI**: Blind users can "feel" any data
+- **Cross-modal learning**: Single representation for all modalities
+
+## Citation
+
+```bibtex
+@misc{braille256v5,
+  title={braille256-v5: Multimodal Universal Braille Model},
+  author={Barrett, Ryan},
+  year={2024},
+  publisher={HuggingFace},
+  url={https://huggingface.co/ryanscottbarrett/braille256-v5}
+}
+```
+
+## License
+
+MIT

config.json

@@ -0,0 +1,9 @@
+{
+  "vocab_size": 32000,
+  "hidden_size": 256,
+  "num_layers": 4,
+  "num_heads": 4,
+  "intermediate_size": 1024,
+  "max_position_embeddings": 512,
+  "dropout": 0.1
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2148874e22f8a0ce4c7ffb4a75bdc5c789440570ea6c3898e72ee334871dd00
+size 45954387

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec5b8b6fbd8985a97c74d377a83f58ff59f3860d02a343eb15146da467da40ae
+size 1155082