Add config.json (architecture metadata, enables Hub download counting); card: download snippet

README.md

@@ -23,14 +23,20 @@ Encoder-decoder transformer trained from scratch for C++ → Python code transla
 
 ## Files
 
+- `config.json` — architecture hyperparameters and vocabulary sizes.
 - `best_model.pt` — full PyTorch checkpoint (model state, optimizer state, src/tgt vocabularies). 189 MB.
 
 ## Load
 
 ```python
-from model import build_transformer
-import torch
-ckpt = torch.load('best_model.pt', map_location='cpu')
+from huggingface_hub import snapshot_download
+repo_dir = snapshot_download("MR0b0t/cpp-to-python-transformer")  # whole repo, incl. config.json
+```
+
+```python
+import os, torch
+from model import build_transformer  # model.py + tokenizer.py from the GitHub repo
+ckpt = torch.load(os.path.join(repo_dir, 'best_model.pt'), map_location='cpu', weights_only=False)
 model = build_transformer(
     src_vocab_size=len(ckpt['src_vocab']),
     tgt_vocab_size=len(ckpt['tgt_vocab']),

config.json

@@ -0,0 +1,19 @@
+{
+  "model_type": "encoder-decoder-transformer",
+  "_note": "Architecture metadata. Also the file the Hugging Face Hub counts downloads on for repos without a registered library.",
+  "d_model": 256,
+  "num_encoder_layers": 4,
+  "num_decoder_layers": 4,
+  "num_heads": 8,
+  "d_ff": 512,
+  "dropout": 0.1,
+  "max_seq_len": 300,
+  "src_vocab_size": 14476,
+  "tgt_vocab_size": 14512,
+  "src_language": "C++",
+  "tgt_language": "Python",
+  "train_data": "XLCoST C++-Python pair_data_tok_full",
+  "checkpoint": "best_model.pt",
+  "epoch": 19,
+  "val_loss": 2.0474
+}