amkdg/ai-cuda-engineer-Embeddings

AI CUDA Engineer Archive Embeddings

Embeddings of SakanaAI/AI-CUDA-Engineer-Archive, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 28,227 AI-generated CUDA kernels → 28,297 vectors
• emb.npy — float16 [28297, 4096]
• meta.parquet — one row per vector, aligned with emb.npy: id, uuid, tag, chunk, n_chunks, count, source_ref
• manifest.json — counts and provenance

Usage

import numpy as np, pyarrow.parquet as pq

emb  = np.load("emb.npy", mmap_mode="r")           # [28297, 4096] float16
meta = pq.read_table("meta.parquet").to_pandas()   # one row per vector, aligned with emb

# A document = consecutive rows sharing one `uuid` (`chunk == 0` marks its start);
# docs longer than 8192 tokens span several chunk-rows.
starts = meta.index[meta.chunk == 0]               # first row of each kernel

Source mapping

Each row carries source_ref = L<level>/T<task>/<kernel_name>|correct=<bool>|speedup_native=<x>, the locator + key metrics back into SakanaAI/AI-CUDA-Engineer-Archive.

Notes

This dataset is not conversational. Each of the 28,227 rows is one AI-generated CUDA kernel composed into a labeled document so the embedding carries both what operation is accelerated and how:

task: <Op_Name>
pytorch_reference: <reference PyTorch op the kernel accelerates>
cuda_kernel: <the generated CUDA code>

All three KernelBench-style levels (level_1/2/3) are combined into one artifact, no dedup (every kernel attempt is a distinct point). tag = Op_Name (the operation family; 229 of them). Correctness and native speedup are packed into source_ref, so correct-vs-broken kernels are filterable straight from meta.parquet (≈54% of archived kernels are correct=True). Rows with empty CUDA_Code were dropped (2,388 of 30,615).