The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
updated: timestamp[s]
note: string
models: list<item: struct<model: string, rsi_C: double, best_of_k: double, self_refine: double, retrieval: d (... 46 chars omitted)
child 0, item: struct<model: string, rsi_C: double, best_of_k: double, self_refine: double, retrieval: double, recu (... 34 chars omitted)
child 0, model: string
child 1, rsi_C: double
child 2, best_of_k: double
child 3, self_refine: double
child 4, retrieval: double
child 5, recursion_gain: double
child 6, budget: int64
split: string
examples: string
reference_code: string
name: string
edge_code: string
id: string
buggy_code: string
tier: string
spec: string
sampler_code: string
fn: string
to
{'name': Value('string'), 'fn': Value('string'), 'spec': Value('string'), 'reference_code': Value('string'), 'buggy_code': Value('string'), 'sampler_code': Value('string'), 'edge_code': Value('string'), 'examples': Value('string'), 'tier': Value('string'), 'split': Value('string'), 'id': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
updated: timestamp[s]
note: string
models: list<item: struct<model: string, rsi_C: double, best_of_k: double, self_refine: double, retrieval: d (... 46 chars omitted)
child 0, item: struct<model: string, rsi_C: double, best_of_k: double, self_refine: double, retrieval: double, recu (... 34 chars omitted)
child 0, model: string
child 1, rsi_C: double
child 2, best_of_k: double
child 3, self_refine: double
child 4, retrieval: double
child 5, recursion_gain: double
child 6, budget: int64
split: string
examples: string
reference_code: string
name: string
edge_code: string
id: string
buggy_code: string
tier: string
spec: string
sampler_code: string
fn: string
to
{'name': Value('string'), 'fn': Value('string'), 'spec': Value('string'), 'reference_code': Value('string'), 'buggy_code': Value('string'), 'sampler_code': Value('string'), 'edge_code': Value('string'), 'examples': Value('string'), 'tier': Value('string'), 'split': Value('string'), 'id': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
KernelAscent — a benchmark for compounding GPU-kernel optimization
KernelAscent asks whether models get better at writing GPU kernels, and whether that improvement compounds recursively. A model writes a kernel; we grade it for correctness against an fp32 reference and for speed against a baseline (eager today; a unified torch.compile baseline reported separately). Scores are headroom-normalized against a per-task roofline, so the ceiling is the model's skill, never the benchmark's.
Site: https://ahmd-mohsin.github.io/KernelAscent/ · Code: https://github.com/ahmd-mohsin/KernelAscent
The five tasks
Tasks 1–4 are building blocks; Task 5 (self-play) is the true recursive-self-improvement metric.
- Capability — one-shot correct+fast kernel. Open and closed models.
- Weight-RSI — an open-weight model LoRA-trains on its own correct kernels; does held-out capability keep rising? (weights are the improvement channel)
- Procedure-RSI — a model rewrites its own executable strategy library + verified archive, weights fixed (works for closed models).
- Closed→Open — a closed frontier model rewrites the training harness of an open trainee; improvement transferred through tooling.
- Self-play (true RSI) — the model authors its own strictly-harder tasks and solves+improves on them, so difficulty and capability co-evolve.
Task 5 — the 3-arm design
Self-play conflates a harder curriculum with an author that co-evolves with the solver. Three arms from the same base, equal budget, one fixed held-out ladder:
- S — STATIC: fixed frontier; solver improves normally.
- F — FROZEN-AUTHOR: frontier escalates, author is the frozen base (adaptive curriculum, non-evolving author).
- L — LIVE-AUTHOR: frontier escalates, author is the current evolving model. Only the author role differs from F.
Decomposition: L−S = total curriculum benefit, F−S = benefit without updating the author, L−F = author co-evolution (the self-referential signal and primary metric). Sustained L−F > 0 with accepted model-authored tasks = genuine recursive compounding; L ≈ F = adaptive-curriculum only. Every authored task passes an anti-reward-hacking gate (rejects constant / identity / no-op / trivial-runtime tasks), is deduplicated, and its provenance is logged. Runs for open-weight (weight channel) and closed API (procedure channel) models.
Mechanism analysis — why RSI fails / how it passes
Alongside the curves we log, per round: generation diversity (mode collapse), predictive entropy, LoRA weight-drift by transformer depth (→0 = ceiling reached), retention on already-solved tasks (forgetting), and the train−held transfer gap (memorization). A verdict attributes each outcome to diversity_collapse | forgetting | drift_saturation | no_transfer | no_headroom, or PASS plus the carrying depth — turning null results (e.g. sub-2B models that never emit a correct kernel) into findings.
Contents
Tiered GPU kernel tasks (DT, class Model(nn.Module), get_inputs()), GPU-validated. Load with:
from datasets import load_dataset
ds = load_dataset("muahmed7338/kernelascent-tasks") # public split, by tier
Each task ships a public development split; official scoring uses a disjoint held-out split.
- Downloads last month
- 336