Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: ValueError
Message: Expected object or value
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 281, in _generate_tables
examples = [ujson_loads(line) for line in batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
episode_id string | scenario string | n_agents int64 | op string | impl string | gpu string | arch string | dtype string | shape unknown | baseline_ms float64 | best_ms float64 | speedup float64 | elapsed_s float64 | agents list | num_messages int64 | messages list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
flash_decode_h100_B=8,H=32,D=128,T=4096,P=16_dispatch_50 | dispatch | 4 | flash_decode | cuda | h100 | sm_90a | fp16 | {
"B": 8,
"H": 32,
"D": 128,
"T": 4096,
"P": 16
} | 0.0336 | 0.011 | 3.057 | 8,506.9 | [
{
"id": "lead",
"role": "lead",
"skill": 1,
"tries": 0,
"kept": 0,
"best_ms": null
},
{
"id": "a01",
"role": "worker",
"skill": 0.8056771987,
"tries": 4,
"kept": 1,
"best_ms": 0.0134694392
},
{
"id": "a02",
"role": "worker",
"skill": 1.0510267877,
... | 50 | [
{
"role": "lead",
"agent": "lead",
"kind": "plan",
"content": "campaign: flash_decode h100 B=8,H=32,D=128,T=4096,P=16. plan: split 3 variants across 4 agents, merge best, redirect losers",
"t": 153.8,
"reward": 0
},
{
"role": "env",
"agent": "env",
"kind": "result",
"cont... |
flash_decode_h100_B=32,H=8,D=128,T=8192,P=16_parallel_sweep_51 | parallel_sweep | 10 | flash_decode | cuda | h100 | sm_90a | fp16 | {
"B": 32,
"H": 8,
"D": 128,
"T": 8192,
"P": 16
} | 0.2319 | 0.0595 | 3.898 | 4,116.1 | [
{
"id": "lead",
"role": "lead",
"skill": 1,
"tries": 0,
"kept": 0,
"best_ms": null
},
{
"id": "a01",
"role": "worker",
"skill": 0.9326706106,
"tries": 2,
"kept": 1,
"best_ms": 0.06603304
},
{
"id": "a02",
"role": "worker",
"skill": 1.0525757228,
... | 51 | [
{
"role": "lead",
"agent": "lead",
"kind": "plan",
"content": "campaign: flash_decode h100 B=32,H=8,D=128,T=8192,P=16. plan: 10-agent parallel cfg sweep on strongest variant, leaderboard merge, top-2 refine",
"t": 140.8,
"reward": 0
},
{
"role": "env",
"agent": "env",
"kind":... |
flash_decode_h100_B=128,H=32,D=64,T=2048,P=32_pipeline_82 | pipeline | 5 | flash_decode | cuda | h100 | sm_90a | fp16 | {
"B": 128,
"H": 32,
"D": 64,
"T": 2048,
"P": 32
} | 0.1183 | 0.0318 | 3.723 | 16,226.7 | [
{
"id": "lead",
"role": "lead",
"skill": 1,
"tries": 0,
"kept": 0,
"best_ms": null
},
{
"id": "a01",
"role": "worker",
"skill": 1.0917963945,
"tries": 11,
"kept": 2,
"best_ms": 0.0317771982
},
{
"id": "a02",
"role": "reviewer",
"skill": 1.007477501... | 82 | [
{
"role": "lead",
"agent": "lead",
"kind": "plan",
"content": "campaign: flash_decode h100 B=128,H=32,D=64,T=2048,P=32. plan: proposer -> review -> verify -> tune pipeline over variants",
"t": 214.2,
"reward": 0
},
{
"role": "env",
"agent": "env",
"kind": "result",
"conte... |
flash_decode_h200_B=8,H=32,D=128,T=4096,P=16_campaign_127 | campaign | 16 | flash_decode | cuda | h200 | sm_90a | fp16 | {
"B": 8,
"H": 32,
"D": 128,
"T": 4096,
"P": 16
} | 0.026 | 0.0088 | 2.969 | 4,062.2 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 127 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode h200 B=8,H=32,D=128,T(...TRUNCATED) |
flash_decode_h200_B=32,H=8,D=128,T=8192,P=16_dispatch_56 | dispatch | 5 | flash_decode | cuda | h200 | sm_90a | fp16 | {
"B": 32,
"H": 8,
"D": 128,
"T": 8192,
"P": 16
} | 0.1647 | 0.0428 | 3.849 | 3,719.2 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 56 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode h200 B=32,H=8,D=128,T(...TRUNCATED) |
flash_decode_h200_B=128,H=32,D=64,T=2048,P=32_parallel_sweep_44 | parallel_sweep | 7 | flash_decode | cuda | h200 | sm_90a | fp16 | {
"B": 128,
"H": 32,
"D": 64,
"T": 2048,
"P": 32
} | 0.0852 | 0.0233 | 3.661 | 8,531.7 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 44 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode h200 B=128,H=32,D=64,(...TRUNCATED) |
flash_decode_a100_B=8,H=32,D=128,T=4096,P=16_pipeline_39 | pipeline | 4 | flash_decode | cuda | a100 | sm_80 | fp16 | {
"B": 8,
"H": 32,
"D": 128,
"T": 4096,
"P": 16
} | 0.0519 | 0.0155 | 3.352 | 10,210.4 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 39 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode a100 B=8,H=32,D=128,T(...TRUNCATED) |
flash_decode_a100_B=32,H=8,D=128,T=8192,P=16_campaign_96 | campaign | 13 | flash_decode | cuda | a100 | sm_80 | fp16 | {
"B": 32,
"H": 8,
"D": 128,
"T": 8192,
"P": 16
} | 0.3775 | 0.0941 | 4.011 | 9,411.3 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 96 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode a100 B=32,H=8,D=128,T(...TRUNCATED) |
flash_decode_a100_B=128,H=32,D=64,T=2048,P=32_dispatch_83 | dispatch | 7 | flash_decode | cuda | a100 | sm_80 | fp16 | {
"B": 128,
"H": 32,
"D": 64,
"T": 2048,
"P": 32
} | 0.1909 | 0.0498 | 3.831 | 17,210.1 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 83 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode a100 B=128,H=32,D=64,(...TRUNCATED) |
flash_decode_a10_B=8,H=32,D=128,T=4096,P=16_parallel_sweep_71 | parallel_sweep | 14 | flash_decode | cuda | a10 | sm_86 | fp16 | {
"B": 8,
"H": 32,
"D": 128,
"T": 4096,
"P": 16
} | 0.1608 | 0.0426 | 3.777 | 7,176.7 | [{"id":"lead","role":"lead","skill":1.0,"tries":0,"kept":0,"best_ms":null},{"id":"a01","role":"worke(...TRUNCATED) | 71 | [{"role":"lead","agent":"lead","kind":"plan","content":"campaign: flash_decode a10 B=8,H=32,D=128,T=(...TRUNCATED) |
End of preview.
kernelswarm — multi-agent kernel-optimization episodes (RL)
Synthetic dataset of multi-agent long-horizon optimization campaigns on AI-inference kernels. Each row is one whole episode: a lead orchestrator plus 2–20 specialist agents split the work, exchange dispatches/statuses/handoffs, submit full kernel candidates, and receive simulated compile/verify/bench feedback — with a reward attached to every environment result.
Inference-focused only. All ops are inference primitives (decode/prefill attention, quantized GEMM/GEMV, sampling, norms, KV-cache, MoE, Mamba).
⚠️ All benchmarks, compiles and correctness checks are simulated analytically — no kernel was ever compiled or executed on a real GPU.
Row format (one row = one episode)
{
"episode_id": "flash_decode_h100_B=8,..._dispatch_74",
"scenario": "dispatch", // dispatch | parallel_sweep | pipeline | campaign
"n_agents": 6,
"op": "flash_decode", "impl": "cuda",
"gpu": "h100", "arch": "sm_90a", "dtype": "fp16",
"shape": {...}, "baseline_ms": 0.42, "best_ms": 0.006,
"speedup": 66.4, "elapsed_s": 17335,
"agents": [{"id": "a01", "role": "worker", "skill": 0.9,
"tries": 8, "kept": 5, "best_ms": 0.008}, ...],
"num_messages": 74,
"messages": [ ... ]
}
Message kinds
| kind | role | meaning |
|---|---|---|
plan |
lead | campaign split strategy |
dispatch |
lead | subtask assignment (to, subtask) |
candidate |
agent | full kernel source + params |
result |
env | compile/verify/bench outcome + reward |
status |
agent | terse progress report |
review |
reviewer | pass/fail verdict |
verify |
verifier | correctness check report |
leaderboard |
env | ranked agent best-times |
merge |
lead | pick winner, update shared best |
redirect |
lead | re-assign agent to winner's variant/region |
handoff |
agent | pass artifact to another agent |
verdict |
lead | final outcome + terminal reward |
Rewards
resultmsgs:min(1, log2(prev_best / new))on kept improvement;-0.05no-improvement,-0.1compile error,-0.3wrong result,-0.05unsupported arch.verdictmsg:log2(speedup) / 4terminal bonus.
Scenarios
- dispatch — lead splits variants across workers, merges, redirects losers (2–8 agents)
- parallel_sweep — N workers each own a cfg slice of one variant; leaderboard + top-2 refine (4–20)
- pipeline — proposer → reviewer → verifier → tuner chain over variants (3–6)
- campaign — fleet dispatch over variant×cfg space + verifier rebench (5–20)
Coverage
- 17 ops (14 CUDA + 3 Triton impls): flash_decode, swa_decode, gemm_fp16, int8_gemm, fp8_gemm, int4_gemv, rmsnorm, rope, swiglu_act, topk_sample, nucleus_sample, mamba_scan, moe_scatter, kv_append, fused_qkv_rope, triton_flash_decode, triton_rmsnorm
- GPUs: h100, h200, a100, a10, rtx_a6000, rtx_5090 (arch features gated: wgmma/tma/fp8/fp4/cp.async)
- 2400 episodes · ~170k messages · avg 71 msgs/episode · avg 8.6 agents
- Simulated horizon: 1–5 h per episode
Files
batches/batch_*.jsonl— 24 shards × 100 episodesepisodes.jsonl— all episodes concatenatedtasks_holdout.jsonl— (op, gpu, shape) task specs for evalstats.json— aggregate stats
- Downloads last month
- 23