dataset

README.md

@@ -27,6 +27,12 @@ ROCmPort AI orchestrates a migration loop:
 
 Complex kernels can fail conversion due to architecture assumptions, undefined behavior, inline PTX, or handcrafted memory logic. The value is reduced migration time and faster debug loops.
 
+## Dataset
+
+Training data: [tazwarrrr/cuda-to-rocm-wavefront-bugs](https://huggingface.co/datasets/tazwarrrr/cuda-to-rocm-wavefront-bugs)  
+170 expert-curated CUDA→ROCm wavefront-64 bug examples across 6 categories.
+See `dataset/upload_dataset.py` for how it was compiled and uploaded.
+
 ## Target User and Business Case
 
 Primary product position:

dataset/upload_dataset.py

@@ -0,0 +1,118 @@
+"""
+Upload cuda-to-rocm-wavefront-bugs dataset to HuggingFace Hub.
+
+Supports either:
+  - 17 individual batch JSON files (recommended)
+  - A single combined JSONL file
+
+Usage (individual files in current dir):
+    python upload_dataset.py --token hf_xxxx --files_dir .
+
+Usage (specific directory):
+    python upload_dataset.py --token hf_xxxx --files_dir ./my_batches/
+
+Usage (single JSONL fallback):
+    python upload_dataset.py --token hf_xxxx --jsonl cuda_rocm_bugs.jsonl
+"""
+import json
+import os
+import glob
+import argparse
+from collections import Counter
+from datasets import Dataset, DatasetDict, Features, Value
+
+FEATURES = Features({
+    "id": Value("string"),
+    "bug_category": Value("string"),
+    "risk_level": Value("string"),
+    "kernel_type": Value("string"),
+    "cuda_snippet": Value("string"),
+    "hip_naive": Value("string"),
+    "hip_corrected": Value("string"),
+    "explanation": Value("string"),
+    "amd_hardware": Value("string"),
+    "rocm_version": Value("string"),
+    "verified_on_mi300x": Value("bool"),
+    "hipify_catches_this": Value("bool"),
+})
+
+
+def load_from_files(files_dir):
+    """Load all batch JSON files from a directory."""
+    pattern = os.path.join(files_dir, "*.json")
+    files = sorted(glob.glob(pattern))
+    if not files:
+        raise FileNotFoundError(f"No .json files found in {files_dir}")
+
+    all_examples = []
+    for f in files:
+        with open(f) as fp:
+            batch = json.load(fp)
+            all_examples.extend(batch)
+        print(f"  Loaded {len(batch):>3} examples from {os.path.basename(f)}")
+
+    return all_examples
+
+
+def load_from_jsonl(jsonl_path):
+    with open(jsonl_path) as f:
+        return [json.loads(line) for line in f]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--token", required=True,
+                        help="HuggingFace write token")
+    parser.add_argument(
+        "--repo", default="tazwarrrr/cuda-to-rocm-wavefront-bugs")
+    parser.add_argument("--files_dir", default=None,
+                        help="Directory with batch .json files")
+    parser.add_argument("--jsonl", default=None,
+                        help="Single combined JSONL file")
+    args = parser.parse_args()
+
+    # Auto-detect if nothing specified
+    if not args.files_dir and not args.jsonl:
+        json_files = sorted(glob.glob("batch_*.json"))
+        if json_files:
+            print(
+                f"Auto-detected {len(json_files)} batch files in current directory")
+            args.files_dir = "."
+        elif os.path.exists("cuda_rocm_bugs.jsonl"):
+            args.jsonl = "cuda_rocm_bugs.jsonl"
+        else:
+            parser.error("Provide --files_dir or --jsonl")
+
+    # Load data
+    print("\nLoading data...")
+    if args.files_dir:
+        data = load_from_files(args.files_dir)
+    else:
+        data = load_from_jsonl(args.jsonl)
+
+    print(f"\nTotal: {len(data)} examples")
+    print("\nBy category:")
+    for cat, count in sorted(Counter(e["bug_category"] for e in data).items()):
+        print(f"  {cat}: {count}")
+
+    # Build dataset with 90/10 split
+    ds = Dataset.from_list(data, features=FEATURES)
+    split = ds.train_test_split(test_size=0.1, seed=42)
+    dataset_dict = DatasetDict(
+        {"train": split["train"], "test": split["test"]})
+
+    print(
+        f"\nSplit: {len(dataset_dict['train'])} train / {len(dataset_dict['test'])} test")
+    print(f"\nUploading to https://huggingface.co/datasets/{args.repo} ...")
+
+    dataset_dict.push_to_hub(args.repo, token=args.token, private=False)
+
+    print("\n✅ Done!")
+    print(f"   https://huggingface.co/datasets/{args.repo}")
+    print("\nNext steps:")
+    print("  1. Paste dataset_README.md as the Dataset Card on HuggingFace")
+    print("  2. Link dataset in lablab.ai submission under 'HuggingFace' track")
+
+
+if __name__ == "__main__":
+    main()