Datasets:
messages listlengths 3 3 | data_source large_stringclasses 5
values | uid large_stringlengths 15 96 | op_type large_stringclasses 1
value | matrix large_stringclasses 510
values | K int64 -1 -1 | run_id large_stringclasses 489
values | iteration int64 1 10 | speedup float64 0 29.9 | n_tokens int64 2 26.4k | enable_thinking bool 2
classes |
|---|---|---|---|---|---|---|---|---|---|---|
[
{
"role": "system",
"content": "You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ source code for kernel.cu — no markdown code fences, no commentary before or after the code.",
"reasoning_content": null
},
{
"role": "user",
"content": "## NCU PROFILING FEEDBACK FROM PREVI... | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_AG-Monien_ukerbe1-20260921-133115__i7 | spgemm | AG-Monien_ukerbe1 | -1 | SpGEMM_AG-Monien_ukerbe1-20260921-133115 | 7 | 3.6377 | 23,076 | true |
[
{
"role": "system",
"content": "You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ source code for kernel.cu — no markdown code fences, no commentary before or after the code.",
"reasoning_content": null
},
{
"role": "user",
"content": "## PERFORMANCE REGRESSION CRITICAL R... | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_vanHeukelum_cage11-20260922-031810__i4 | spgemm | vanHeukelum_cage11 | -1 | SpGEMM_vanHeukelum_cage11-20260922-031810 | 4 | 0.8844 | 21,947 | true |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_Sanghavi_ecl32-20260922-011902__i8 | spgemm | Sanghavi_ecl32 | -1 | SpGEMM_Sanghavi_ecl32-20260922-011902 | 8 | 1.1852 | 21,369 | true |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_DIMACS10_fe_rotor-20260921-210513__i6 | spgemm | DIMACS10_fe_rotor | -1 | SpGEMM_DIMACS10_fe_rotor-20260921-210513 | 6 | 0.5773 | 21,558 | false |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_JGD_Kocay_Trec13-20260921-130658__i3 | spgemm | JGD_Kocay_Trec13 | -1 | SpGEMM_JGD_Kocay_Trec13-20260921-130658 | 3 | 1.169 | 17,204 | false |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_LiuWenzhuo_powersim-20260921-163018__i4 | spgemm | LiuWenzhuo_powersim | -1 | SpGEMM_LiuWenzhuo_powersim-20260921-163018 | 4 | 1.2579 | 19,915 | false |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_DIMACS10_rgg_n_2_16_s0-20260921-213154__i6 | spgemm | DIMACS10_rgg_n_2_16_s0 | -1 | SpGEMM_DIMACS10_rgg_n_2_16_s0-20260921-213154 | 6 | 0.7519 | 20,786 | true |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_JGD_Trefethen_Trefethen_2000-20260921-131212__i4 | spgemm | JGD_Trefethen_Trefethen_2000 | -1 | SpGEMM_JGD_Trefethen_Trefethen_2000-20260921-131212 | 4 | 3.1753 | 18,922 | false |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_JGD_GL7d_GL7d13-20260921-163018__i4 | spgemm | JGD_GL7d_GL7d13 | -1 | SpGEMM_JGD_GL7d_GL7d13-20260921-163018 | 4 | 0.9131 | 20,593 | true |
[{"role":"system","content":"You are a CUDA sparse kernel specialist. Output ONLY the raw CUDA/C++ s(...TRUNCATED) | drsparse_gpt56_distill | luna_pool562_single10__SpGEMM_Mittelmann_watson_1-20260921-225938__i7 | spgemm | Mittelmann_watson_1 | -1 | SpGEMM_Mittelmann_watson_1-20260921-225938 | 7 | 1.2368 | 20,987 | false |
End of preview. Expand in Data Studio
Dr.Sparse SFT data — Luna-10 distillation, round v4b (Qwen3.5-9B student)
The exact data behind the qwen3.5-9b-sft-v4b student (OTF-81: 46 % of held-out matrices solved with
reasoning on, 58 % with reasoning off; the base model solves 0). Built from GPT-5.6 ("Luna")
trajectories on the 562-matrix Dr.Sparse training pool; no OTF test-set matrix appears anywhere.
| path | what | rows |
|---|---|---|
qwen3.5-9b-v4b/{train,val}.parquet |
train on this. Rendered for the Qwen3.5 chat template: messages (system = eval coder prompt, user = coder task prompt, assistant = kernel, reasoning_content = teacher-written trace on 2,134 rows), enable_thinking, matrix, speedup, provenance columns |
3,963 / 71 |
qwen3.5-9b-mixed/{train,val}.parquet |
the v3 "mixed" parquet (each Luna kernel once plain, once with a Qwen3-32B rationale); input of star_to_sft.py |
2,489 / 71 |
sources/luna10_all_turns.jsonl |
every coder turn of the Luna single-chain 10-round run over 562 matrices (5,610 turns, with harness results) | |
sources/luna_rs_round1/ |
teacher rejection sampling on iteration-1 prompts (K=4 + 1 repair), with re-benchmarked speedups | 2,192 |
sources/rationales_luna.jsonl |
teacher-written <think> traces for the kernels (rationalization protocol) |
3,329 |
sources/star_round1/ |
student (Qwen3.5-9B v3) self-samples and rationalizations | |
sources/rationales_qwen3_32b.jsonl |
Qwen3-32B rationales used by the mixed parquet | |
passk_prompts/ |
the 10-prompt pass@K coverage probe (training/distill/passk_eval.py) |
10 |
Composition of qwen3.5-9b-v4b/train.parquet by data_source: Luna 10-round kernels 1,312; the same
kernels with a Luna-written trace 1,270; Luna rejection-sampling kernels 905; student rationalizations
430; student self-samples 46. All kernels compiled, passed memcheck and matched cuSPARSE (any speed).
Code, recipe and the one-command reproduction: training/distill/RUNBOOK.md in the Dr.Sparse repo
(branch v2), script training/distill/run_v4b_sft.sh.
- Downloads last month
- 107