Genghan/PTXBench-Qwen3.6-27B-s0
Image-Text-to-Text • Updated • 5
id stringlengths 37 50 | messages listlengths 3 3 | metadata dict |
|---|---|---|
2026-0503-2313_exp_002_mha_with_lse_h48_d128_t5 | [
{
"content": "You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel using TVM-FFI binding.\n\n## Response Format\n\nEach response, provide your complete kernel implementation inside a single ```cpp code block.\nYou will receive evaluation feedback showing compilation errors, co... | {
"arch": "hopper",
"arch_tag": "H",
"assistant_target_source": "trajectory_actual_gemini_output",
"base_prompt_gpu_arch": "hopper",
"base_prompt_sha256": "88f8afa7cab4f8f51227720070b7b03ab7f2d40fbc8f93be9a984081a23127fe",
"base_prompt_sources": [
"/home/ubuntu/AccRL/fib_runtime/structural_doc/document/... |
2026-0503-2313_exp_003_mha_with_lse_h48_d128_t2 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_000_mha_with_lse_h48_d128_t2 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_001_mha_with_lse_h48_d128_t2 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_003_mha_with_lse_h48_d128_t3 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_001_mha_with_lse_h48_d128_t5 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_000_mha_with_lse_h48_d128_t7 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_000_mha_with_lse_h48_d128_t6 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_003_mha_with_lse_h48_d128_t5 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
2026-0503-2313_exp_004_mha_with_lse_h48_d128_t5 | [{"content":"You are an expert CUDA kernel developer. Your task is to write an optimized CUDA kernel(...TRUNCATED) | {"arch":"hopper","arch_tag":"H","assistant_target_source":"trajectory_actual_gemini_output","base_pr(...TRUNCATED) |
End of preview. Expand in Data Studio
PTXBench Qwen3.6-27B SFT datasets
This private repository contains the byte-exact parquet files used to train PTXBench Qwen3.6-27B-s0 through Qwen3.6-27B-s6. Load one release as:
from datasets import load_dataset
dataset = load_dataset("Genghan/PTXBench-Qwen3.6-27B-SFT", "s0", split="train", token=True)
| Config | Internal recipe | Scope | Template | Reasoning synthesizer | Rows | SHA-256 |
|---|---|---|---|---|---|---|
| s0 | sft-v4 |
8ops-Extended | KernelGen | GLM-5.2 | 494 | 5416899bf9f8312e1e5361dc12f20613246ef597f073852224c81eb314c10ff8 |
| s1 | fixit-v2-glm |
4ops | Fixit | GLM-5.2 | 158 | 6a42a93125cc8c6dfc0bb52ec1c807e0e6b39fd8f04c43300d82f82ad866d4de |
| s2 | fixit-v2-glm-8turns |
4ops-Extended | Fixit | GLM-5.2 | 259 | 06d78bfa4c6f68ec73d7f8d57df4eb31c2a946d1915edfa1c19202e7b74448b1 |
| s3 | fixit-v4 |
8ops-Extended | Fixit | GLM-5.2 | 406 | 8aa2bf4f4c34e542bcc13c66f16e18f51d0da20426823108c8d9778c0002b0b8 |
| s4 | fixit-v5 |
8ops-Post-balanced | Fixit | GLM-5.2 | 170 | eeda2fa32ded94675fa8b800b3a7be110d79eec195945e836d9141c11f4b3664 |
| s5 | fixit-v5-full |
8ops-Pre-balanced | Fixit | GLM-5.2 | 258 | e59150ed81e28edf92710f06760a77077d54d588a3b85ec8d641d52dca5f33d5 |
| s6 | fixit-v6 |
8ops-Pre-balanced | Fixit | Qwen3.6-27B | 258 | 55122991454816ab51ffbc10d5be2547bb7000b44e8f4df93aa7804a67a60689 |
The parquet metadata column retains historical machine-local paths because
these are the exact training artifacts. Those paths are provenance strings and
are not required to load or train on the messages column.
The s0 rows contain system, user, and assistant messages. The s1-s6
Fixit rows contain system, user, assistant, user, and assistant
messages, with only the final assistant message carrying loss mask 1.
All seven runs used Qwen/Qwen3.6-27B, five epochs, learning rate 4.65e-4,
maximum length 65,536, and LoRA rank 32. See manifest.json for exact
artifact and checkpoint provenance.
These data contain generated reasoning and CUDA kernels. Repository access is private; no public redistribution license is asserted by this card.
- Downloads last month
- 12