Zhangdanyang/final-hip-gfx950-all-v2

GEAK gfx950 Strict SFT Dataset

• Backend: hip
• Architecture: gfx950
• Samples: 244
• Quota trimmed: false
• Deduplication: contract_hash + patch_hash
• Task modes: {"cold_start": 30, "direction_conditioned": 80, "error_recovery": 25, "profile_guided": 92, "regression_balance": 17}
• Split scope: train-only; dev and held-out are deferred and not inferred by ETL.
• Leakage report scope: conflicts among explicit split metadata present in samples.

Kernels by operator

Case kernels are distinct contract/harness entries. Candidate kernels are verified optimized SFT samples.

• knn: 30 case kernels; 55 candidate kernels
• rms_norm: 15 case kernels; 42 candidate kernels
• silu_and_mul: 13 case kernels; 47 candidate kernels
• gemm: 12 case kernels; 34 candidate kernels
• softmax: 9 case kernels; 29 candidate kernels
• fused_add_rms_norm: 1 case kernels; 31 candidate kernels
• fused_moe: 1 case kernels; 1 candidate kernels
• mha: 1 case kernels; 1 candidate kernels
• mla_decode: 1 case kernels; 3 candidate kernels
• paged_attention_decode: 1 case kernels; 1 candidate kernels

Load from a downloaded repository

import json
from pathlib import Path

root = Path('.')
sample = json.loads((root / 'samples.jsonl').open().readline())
parent_code = (root / sample['parent_sources']['kernel.py']['path']).read_text()
candidate_code = (root / sample['candidate_sources']['kernel.py']['path']).read_text()
candidate_workspace = root / sample['artifact_paths']['candidate_workspace']

Each parent and candidate workspace includes config.yaml, metadata.json, and scripts/task_runner.py for local replay. All functional paths are relative to the dataset root.